Free skills that read the UN source first
Each one is a single file. Add it to Claude and it reads the official UN source before it answers. If it cannot read the source, it says so instead of guessing.
btr-checklist
Build a preparation checklist for a country's next Biennial Transparency Report (BTR) under the Enhanced Transparency Framework of the Paris Agreement, with every item and every date traced to a UNFCCC decision paragraph. Use when someone asks what a country must submit in its BTR, by when, and what flexibility or discretion applies.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: btr-checklist description: Build a preparation checklist for a country's next Biennial Transparency Report (BTR) under the Enhanced Transparency Framework of the Paris Agreement, with every item and every date traced to a UNFCCC decision paragraph. Use when someone asks what a country must submit in its BTR, by when, and what flexibility or discretion applies. --- # BTR Checklist One country situation in, one checklist out. Each line names the decision and paragraph it comes from. A date that no opened document states is marked unverified. **Inputs** - `country`: optional. Matters only for whether it is a least developed country (LDC) or small island developing State (SIDS), and whether it is a developing country Party applying flexibility. - `report`: which BTR, e.g. the second. ## Why this exists BTR questions get answered from slides and blog posts that mix the old biennial update reports with the new framework, and that state deadlines without saying where they come from. A deadline in a brief needs a decision number and a paragraph, or it needs the word unverified next to it. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | Decision 1/CP.21, FCCC/CP/2015/10/Add.1 | Para 90, how often Parties report under Article 13 | `https://unfccc.int/resource/docs/2015/cop21/eng/10a01.pdf` | | Decision 18/CMA.1 and its annex (the MPGs), FCCC/PA/CMA/2018/3/Add.2 | Decision para 3 (first BTR date), para 4 (LDC and SIDS discretion). MPG paras 5 and 6 (flexibility), para 10 (what a BTR contains), paras 11 to 16 (format and submission), para 58 (latest inventory year), para 198 (late submission) | `https://unfccc.int/sites/default/files/resource/CMA2018_03a02E.pdf` | | UNFCCC page for the report in question | The secretariat's statement of the due date and the list of submissions | `https://unfccc.int/second-biennial-transparency-reports` (for BTR2) | | Paris Agreement, Article 13 | The paragraphs 7 to 10 the decisions refer to. The Agreement is the annex to decision 1/CP.21 | Same file as decision 1/CP.21 | The UNFCCC site often serves a bot check instead of the page. The PDFs usually open with a plain download. If a page does not open, say so and do not fill it from memory or a search summary. ## Steps 1. Open both decision PDFs. Copy each document symbol and its date from the first page. 2. Find the due date. Quote decision 18/CMA.1 para 3 for the first BTR and decision 1/CP.21 para 90 for the frequency. 3. For a later BTR, open the UNFCCC page for that report and quote its date. If the page does not open, write the date as derived from the two paragraphs above and mark it `UNVERIFIED: not stated in an opened document`. 4. Build the checklist from MPG para 10, one line per sub-paragraph, keeping the decision's own verb. "Shall" is required. "Should" is expected. Do not turn a "should" into a must. 5. Add the format and submission lines from MPG paras 12, 15 and 16, the inventory year rule from para 58, and the LDC and SIDS line from decision para 4 and MPG para 11. 6. If the country is a developing country Party, add the flexibility rule from MPG para 6 in one line. ## Answer layout ``` <report> · sources read <today's date> Due: <date> "<quoted line>" <decision>, para <n> (or UNVERIFIED and why) Frequency: "<quoted line>" decision 1/CP.21, para 90 | # | Item | Shall or should | Source | |---|---|---|---| | 1 | <item in the decision's words> | shall | MPG para 10(a) | | ... | | | | Not covered: <anything asked about that the opened documents do not state> ``` ## Rules - **A date with no paragraph behind it is `UNVERIFIED`. Never give a deadline from memory, a news item or a search result summary as confirmed.** - Quote the decision's words for every item. Shorten only by cutting, never by paraphrase that changes "shall" or "should". - Keep BTRs apart from national communications and from biennial update reports. They are different reports under different decisions. - If the question asks about a tool, a template or a table the opened documents do not describe, write `NOT IN SOURCE` and name the document that would hold it, if known. - Never invent a number of tables, pages or chapters. ## Copy-paste prompt (works in any assistant) > Make me a checklist for <country>'s <first/second> Biennial Transparency Report under the Paris Agreement. Use only UNFCCC decisions: decision 1/CP.21 paragraph 90, decision 18/CMA.1 and its annex of modalities, procedures and guidelines (paragraph 10 for contents), and the UNFCCC page for that report. For each item, give the decision paragraph and say whether the text says "shall" or "should". For the due date, quote the paragraph or page that states it. If no document you opened states the date, write UNVERIFIED next to it. Do not answer from memory. ## Running it as code Copies of the two decision PDFs used for the recorded test, with SHA-256, are in `data/anchors/skills/btr-checklist/` in the claude-for-sdg repository. The UNFCCC page for BTR2 could not be opened when this skill was written.
claim-check
Check a short draft (a brief, talking points, a donor note) sentence by sentence against the public UN documents it cites. Returns one table row per sentence with the source, the page or section, the line quoted word for word, and a verdict of SUPPORTED, NOT IN SOURCES or WRONG PAGE. Use before a draft with figures goes to a reviewer, a donor or a meeting.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: claim-check description: Check a short draft (a brief, talking points, a donor note) sentence by sentence against the public UN documents it cites. Returns one table row per sentence with the source, the page or section, the line quoted word for word, and a verdict of SUPPORTED, NOT IN SOURCES or WRONG PAGE. Use before a draft with figures goes to a reviewer, a donor or a meeting. --- # Claim Check A draft and its sources in, one row per sentence out. A sentence passes only when a line in the source says the same thing on the page the draft cites. **Inputs** - `draft`: up to about 15 sentences. Number them if they are not numbered. - `sources`: the documents the draft cites, as links or files. Official UN documents only (see below). ## Why this exists A reviewer checks the number and the page. A checker that reads from memory, or matches only the numbers, passes a sentence whose figure sits on another page or belongs to another sentence. Long UN reports repeat the same percentages for different subjects, so a number found on the right page is not yet a match. ## Sources (use these and nothing else) - The documents the user gives, if they are published by the UN system: documents with a UN symbol (A/..., E/..., FCCC/...) from `documents.un.org` or the publishing office, and reports on `un.org`, `unstats.un.org` or a UN agency's own site. - Nothing else. Do not confirm a sentence from memory, from a news story or from a different edition of the same report. - A source the user names but you cannot open is not a source. Every sentence that cites it gets `NOT IN SOURCES` with the note "could not open". ## Steps 1. Open each source. Copy its symbol or title and its date as printed on the first page. 2. Work out how the document numbers its pages. Use the page number printed on the page. Say so if it differs from the PDF page count. 3. For each sentence, go to the page or section it cites. Look for a line that states the same claim: the same figure, the same year, the same unit and the same subject. 4. If the cited page has no such line, search the whole source. If the line is on another page, the verdict is `WRONG PAGE` and the row gives the right page. 5. If no page in any source has the line, the verdict is `NOT IN SOURCES`. Quote the closest line, if there is one, so the reader sees what the source does say. 6. Fill one row per sentence. Never merge two sentences into one row. ## Verdicts | Verdict | Means | |---|---| | SUPPORTED | A line on the cited page states the same figure, year, unit and subject | | WRONG PAGE | The line exists in the source but not on the page the draft cites | | NOT IN SOURCES | No line in any source states it. This includes a figure that differs from the source by any amount | A number that matches but belongs to a different subject is not support. A share of adults with a bank account does not support the same share of people with electricity. ## Answer layout ``` Source: <symbol or title>, <date as printed> · <URL> · read <today's date> Pages: <printed page numbers | PDF page numbers> | # | Sentence | Source | Page/section | Quoted line | Verdict | |---|---|---|---|---|---| | 1 | <first words of the sentence> | <symbol> | p. <n>, para <n> | "<line, word for word>" | SUPPORTED | | 2 | ... | ... | cited p. <n>, found p. <m> | "<line>" | WRONG PAGE | | 3 | ... | ... | cited p. <n> | closest: "<line>" or none | NOT IN SOURCES | <n> of <total> supported. ``` ## Rules - **Cannot open a source: say which one and mark its sentences NOT IN SOURCES with "could not open". Do not judge them from memory.** - Quotes are word for word. If text extraction breaks spacing, fix only the spacing. - Never round, convert or recompute a figure to make it match. A source that says 12% does not support 13%. - Do not rewrite the draft. The table says what is wrong, the author fixes it. - Do not add sources the draft did not cite. If a better source exists, say so in one line under the table. ## Copy-paste prompt (works in any assistant) > Check my draft against its sources, one sentence at a time. For each sentence, open the source and the page it cites, and give me a table: sentence, source, page or section, the line from the source word for word, and a verdict. The verdict is SUPPORTED only if a line on the cited page states the same figure, year and subject. If the line is on another page, write WRONG PAGE and give the right page. If no line in the source says it, write NOT IN SOURCES and quote the closest line. Do not judge any sentence from memory. If you can't open a source, say so for every sentence that cites it. ## Demo `demo/draft_brief.md` is a five-sentence draft citing the Secretary-General's SDG progress report A/81/93-E/2026/73 (8 June 2026). Two sentences carry planted errors, and the answer key is in the same file. A copy of the report and its SHA-256 are in `data/anchors/skills/claim-check/` in the claude-for-sdg repository.
classified-intake-steps
Turn a classified item that arrived by email into a short intake card for UN staff. The card says to print it on receipt, file and store it as the bulletin directs, keep it on protected systems, and use AI only on the unclassified parts of the work. Every step is quoted from the Secretary-General's bulletin on classified information. Use when someone has a classified attachment in their inbox and asks whether AI can summarise it.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: classified-intake-steps description: Turn a classified item that arrived by email into a short intake card for UN staff. The card says to print it on receipt, file and store it as the bulletin directs, keep it on protected systems, and use AI only on the unclassified parts of the work. Every step is quoted from the Secretary-General's bulletin on classified information. Use when someone has a classified attachment in their inbox and asks whether AI can summarise it. --- # Classified Intake Steps One classified item in, one intake card out. Each step quotes the bulletin. Anything the opened text does not state is marked NOT IN SOURCE. **Inputs** - `document`: the user's own working notes about the item. This can be the cover email with the attachment removed, or a short note of what arrived and how it is marked. It is never the classified content itself. - `question`: what the user wants to do with the item, such as get a summary or pass it to a colleague. ## Why this exists A classified attachment looks like any other file in an inbox. The quick move is to drop it into an AI tool for a summary. The bulletin asks for a paper step on receipt and protected systems for any electronic handling. Unless your office has approved an AI tool for classified work, treat it as an unprotected system. Staff need the steps at the moment the file arrives, with the bulletin's own words behind each one. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | Secretary-General's bulletin on classified information | The paragraph on printing items received in electronic form, the paragraph on filing and storage it points to, the paragraph on protected means of communication, and the paragraph on controls for automated information systems | `https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf` | | The user's own document | What arrived, how it is marked, and what the user wants to do | Pasted or attached by the user | The UN documents site sometimes returns an error or a bot check instead of the PDF. If the bulletin does not open, say so plainly. Do not fill the steps from memory or a search summary. Give the user the link and stop. ## Steps - Read the user's document first. If it holds classified content, stop. Tell the user to remove it from the chat and not to paste it again. Work only from the marking and the user's own notes. - Open the bulletin PDF. Copy the document symbol and the title from the first page. - Find the paragraph that begins "(e) A hard copy of classified information received in an electronic form". Quote it as the receipt step. - Find the subparagraph that the receipt step points to with the words "as detailed in subparagraph (d) above". Quote what it says on filing and storage. If the text is not there, write NOT IN SOURCE. - Find the paragraph that begins "(f) Electronic transmission". Quote it as the sending step. - Find the paragraph that begins "Heads of departments and offices". Quote the part on "controls that both prevent access by unauthorized persons". - Answer the user's question. If they asked AI to summarise, read, translate or rewrite the classified item, say no and quote the sending step. Then list what AI can still help with on the unclassified side. Examples are a reply to the sender asking which protected channel to use, a plan for the user's own unclassified tasks, or a checklist built from this card. - Say who decides which systems are protected. The bulletin gives this to heads of departments and offices working with IT services. Tell the user to ask their office which systems are approved. Do not name a system yourself. ## Answer layout ``` Intake card · <document symbol from the first page> Item: <what arrived and its marking, from the user's notes> On receipt: print it. "<quoted line>" <paragraph label> Store it: as the bulletin says. "<quoted line>" <paragraph label> (or NOT IN SOURCE) Sending it: protected means only. "<quoted line>" <paragraph label> Systems: approved controls only. "<quoted line>" <paragraph label> AI on this item: no. Use it only for the unclassified parts below. - <unclassified task AI can help with> - ... Ask your office: <which protected systems are approved, and anything else the bulletin leaves to the office> Not covered: <anything asked that the bulletin does not state> ``` ## Rules - **Remove names and personal details first. Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows.** - Never summarise, translate, rewrite or analyse the classified content. If it is pasted, stop and say so. - Quote the bulletin for every step. Shorten only by cutting. Keep "must" and "shall" as written. - If the bulletin does not open, say it could not be opened. Do not fill the steps from memory, a news item or a search summary. - Never say that a named AI tool or platform is approved for classified information. The bulletin leaves approval to heads of departments and offices. - If the question asks about a classification level, a form or a procedure the opened text does not describe, write NOT IN SOURCE and point the user to their office. - Never invent a paragraph label, a classification level or a contact. ## Copy-paste prompt (works in any assistant) > I have a classified email attachment and I want to know what to do with it. Do not ask me to paste it and do not summarise it. Use only the Secretary-General's bulletin at https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf and my own notes below. Make me a short intake card. Quote the bulletin's lines on printing classified items received in electronic form, filing and storing them, sending them only through protected means of communication, and controls on automated information systems. Give the paragraph label for each quote. Then list what AI can help me with on the unclassified parts of my work only. If you cannot open the bulletin, say so and do not answer from memory. My notes: <what arrived, how it is marked, what I want to do>.
country-climate-card
Make a one-page climate card for any country from three open datasets, with every number computed in code. Use when someone preparing a bilateral meeting, a briefing or a funding proposal gives a country name and wants its capital's heat trend, its emissions and biggest source, and its latest NDC on one page.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: country-climate-card description: Make a one-page climate card for any country from three open datasets, with every number computed in code. Use when someone preparing a bilateral meeting, a briefing or a funding proposal gives a country name and wants its capital's heat trend, its emissions and biggest source, and its latest NDC on one page. --- # Country climate card One input: a country name. One output: a card a delegation adviser can read in ten seconds before a bilateral. Every number comes from an API response you saved. Nothing comes from memory. ## 1. Pull the data (save each response to a file before you read it) Use the ISO3 code (Kenya = `KEN`) and the capital's latitude and longitude to two decimals. | What | URL | |---|---| | Daily max temperature, capital | `https://archive-api.open-meteo.com/v1/archive?latitude=<lat>&longitude=<lon>&start_date=1961-01-01&end_date=2025-12-31&daily=temperature_2m_max&timezone=auto` | | Total emissions, one call per year 2015 to 2024 | `https://api.climatetrace.org/v6/country/emissions?countries=<ISO3>&since=<year>&to=<year>` | | Emissions by sector, 2024, one call per sector | same URL with `&since=2024&to=2024§ors=<sector>` for `agriculture buildings fluorinated-gases forestry-and-land-use fossil-fuel-operations manufacturing mineral-extraction power transportation waste` | | NDC content | `https://www.climatewatchdata.org/api/v1/data/ndc_content?countries[]=<ISO3>&per_page=500` | Climate TRACE: read `[0].emissions.co2e_100yr` (tonnes). `since` and `to` must be the same year, or you get a sum of years. Open-Meteo is CC BY 4.0: credit it on the card. ## 2. Compute in code (Python, standard library is enough) - **Hot-day line** `T` = the 90th percentile of all daily maxima 1961–1990, rounded to a whole degree. - **Hot days per year** = days with max ≥ `T`. `base` = mean 1961–1990, `recent` = mean 2016–2025, `ratio` = recent / base. - **Emissions**: total 2024 and change since 2015, in Mt CO₂e. - **Sector shares**: drop `forestry-and-land-use` (it is often a net sink and swamps the rest), then each sector's share of the sum of the others. `top` = largest. - **NDC**: for `ghg_target`, `conditionality` and `submission`, take the row with the highest `id` (the latest submission). - **The sentence** (same rule for every country): - power share < 10% and top is not power → «Power is not its problem. Its <top> are.» (a small power share is not a clean grid, so never write «clean») - otherwise → «Its <top> lead, at <share>%. <Second> come next, at <share>%.» - nouns: agriculture farms, transportation roads, manufacturing factories, buildings buildings, waste landfills, fossil-fuel-operations oil and gas fields, mineral-extraction mines, fluorinated-gases cooling gases. ## 3. Lay out the card (4:5, 1080 × 1350) Top to bottom: the prompt line · country name · **hero `<ratio>×` more hot days in <capital>** with base and recent under it · bar chart of hot days per year 1961–2025, 2016–2025 in a warm colour, flat lines at base and recent · the sentence, second line bold · three tiles (emissions 2024 · biggest source · latest NDC) · one source line. Phone rule: nothing a reader must read under 24 px. ## City mode («heat card for <city>») Open-Meteo only, with `daily=temperature_2m_max,temperature_2m_min` at the city's coordinates. Same hot-day line `T`. Hero = `recent − base` more hot days a year. Add warm nights: nights with minimum above 20 °C, 1961–90 mean against 2016–25 mean. Tiles: the hot-day line `T`, the worst year (most hot days), the hottest day in the series. ## 4. Hand back The card, then a JSON block with `T base recent ratio total24 growth top share_top power_share ndc_submission conditionality sentence`, then the raw `ghg_target` sentence word for word. ## What a person still checks - The NDC line against the country's own NDC document on unfccc.int. Climate Watch is a reading of it. - The grid cell: ERA5 is about 25 km. A coastal capital can land on sea. - Emissions are Climate TRACE estimates, not the national inventory. Say so if the card goes into an official brief.
declassification-sequence-check
Check the order of steps in an office's own plan for sending closed classified files to the Archives and Records Management Section, against the Secretary-General's bulletin on information sensitivity, classification and handling. Use when someone is packing project files for the Archives, or has a transfer plan, and asks whether declassification review has to happen first.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: declassification-sequence-check description: Check the order of steps in an office's own plan for sending closed classified files to the Archives and Records Management Section, against the Secretary-General's bulletin on information sensitivity, classification and handling. Use when someone is packing project files for the Archives, or has a transfer plan, and asks whether declassification review has to happen first. --- # Declassification Sequence Check You give it one transfer plan and get back one sequence card, with each step tied to a quoted line of the bulletin. **Inputs** - `document`: required. The user's own transfer plan, checklist, email or file list. Working notes are fine. - `markings`: optional. The classification levels the files carry, in the user's own words. ## Why this exists Offices close a project, pack the boxes and send them to the Archives. Often nobody thinks of declassification review until the boxes have gone. The bulletin sets the order. Review comes first. Transfer to the Archives comes after. A plan with these steps the wrong way round needs to be caught before the boxes move. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | Secretary-General's bulletin on information sensitivity, classification and handling | The declassification section. The paragraph on marking a date or event that triggers declassification. The paragraph on review before transfer to the Archives. The sub-paragraph on item-by-item review of strictly confidential records | `https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf` | | The user's own document | The steps the office plans to take, in its own order | Pasted or attached by the user | If the bulletin does not open, say so in the first line of the answer. Do not fill in its wording from memory or from a search summary. Check the user's document for order only, and mark every bulletin line UNVERIFIED. ## Steps - Open the bulletin. Copy its symbol and title from the first page. - Find the declassification section. Copy the paragraph that begins "4.1 The originator of the information". - Copy the paragraph that begins "4.3 Review for possible declassification shall take place before records are transferred". - If the user's document mentions strictly confidential files, copy the sub-paragraph that begins "(a) Records that are classified as “strictly confidential”". - Read the user's document. List each step in the order it appears. Keep the user's words. - Put each step into one of three stages. Set a declassification date or event. Review the records. Transfer them to the Archives and Records Management Section. Steps that fit none of these go under Other. - Compare the user's order with the bulletin's order. Flag any transfer that comes before review. Flag any classified file that has no date or event marked on it. Keep the bulletin's words "where appropriate" next to that flag. - Write the sequence card. ## Answer layout ``` Sequence check · bulletin read <today's date> (or: bulletin did not open, every quote UNVERIFIED) Your order: <step> / <step> / <step> Bulletin order: Set: "<quote from the marking paragraph>" Review: "<quote from the review before transfer paragraph>" Transfer: to the Archives and Records Management Section, after review | Your step | Stage | In order? | Bulletin line | |---|---|---|---| | <user's words> | review | no, it comes after transfer | "<quote>" | Strictly confidential: "<quote from sub-paragraph (a)>" (only if the document mentions it) Fix: <which step to move, in one line> Not covered: <anything asked that the bulletin does not state> ``` ## Rules - **Review comes before transfer. Quote the bulletin for this every time. Never state the order from memory.** - Quote the bulletin's own words. Shorten only by cutting. Keep "shall" and "where appropriate" exactly as written. - Work only on the user's document. Do not add steps it does not mention, apart from the three stages. - Do not name who in the user's office may carry out the review. Quote what the bulletin says about who reviews and stop there. - How to recall boxes that have already gone is NOT IN SOURCE. Say so. Name the Archives and Records Management Section as the office that holds them. - Do not work out when a record becomes old enough for review. Quote the sub-paragraph and leave the counting to the user. - Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. ## Copy-paste prompt (works in any assistant) > Check the order of steps in my office's plan for sending closed project files to the UN Archives. My plan is pasted below. Use only the Secretary-General's bulletin on information sensitivity, classification and handling at https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf and my own plan. Quote the declassification paragraph on marking a date or event, the paragraph on review before transfer to the Archives and Records Management Section, and, if my plan mentions strictly confidential files, the sub-paragraph on item-by-item review. List my steps in my order, say which stage each one belongs to, and flag any transfer that comes before review. If you cannot open the bulletin, say so and mark every quote UNVERIFIED. Do not answer from memory.
distribution-list-audit
Check the user's own distribution list or copy log for a recurring classified product, such as a situation report, against the UN Secretary-General's bulletin on information sensitivity, classification and handling. Flags copies sent outside the standard list, copies made without authorization, and copies missing from the special registry. Use when someone asks whether a forward or an extra copy was allowed, or wants an auditable distribution list and copy log.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: distribution-list-audit description: Check the user's own distribution list or copy log for a recurring classified product, such as a situation report, against the UN Secretary-General's bulletin on information sensitivity, classification and handling. Flags copies sent outside the standard list, copies made without authorization, and copies missing from the special registry. Use when someone asks whether a forward or an extra copy was allowed, or wants an auditable distribution list and copy log. --- # Distribution List Audit One distribution list or copy log goes in and one audit comes out. Each finding quotes the line of the bulletin it rests on. The audit holds roles and document IDs only. It never holds the content of the report. **Inputs** - `document`: the user's own distribution list, copy log, or short note of who received which product. Replace names with roles before pasting. - `product`: optional. The recurring product in question, for example a situation report, an operational update or a periodic political assessment. ## Why this exists Recurring reports get forwarded on request. A colleague asks, the report looks useful, and the copy never reaches the registry. The bulletin treats this as a gap in control. Copying a classified sitrep needs authorization. Every copy goes in the special registry. Staff need a quick check against the actual text and a list format that an auditor can follow. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | Secretary-General's bulletin "ST/SGB/2007/6" on information sensitivity, classification and handling | The section on handling classified information. It covers the special registry, the rule on duplication, and standard distribution lists for recurrent products | `https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf` | | The user's own document | The list, the log, or the note of who received what | Supplied by the user | The UN documents site sometimes fails or returns an empty page. If the bulletin does not open, say so plainly. Do not fill the rules from memory or from a search summary. Give the audit layout with each rule marked `NOT CHECKED: source did not open`. ## Steps - Open the bulletin. Copy its symbol and title from the first page. - Find and quote exactly the lines this skill rests on. These are the line on standard distribution lists for recurrent products, the line on the special registry, and the line on duplication. - Read the user's document. List each recipient role and each document ID. Do not repeat any report content. - Mark each recipient as on the standard list or off it. - For each copy, check whether the document records who authorized it. Only the originator or the head of the receiving or originating department or office can authorize a copy. - For each copy, check whether the document shows a registry entry. - Build the clean template with a standard list, a copy log and the open gaps. ## Answer layout ``` <product> · bulletin opened <today's date> (or NOT OPENED and why) Rules checked Standard list: "<quoted line>" Registry: "<quoted line>" Copies: "<quoted line>" | Recipient role | Document ID | On standard list | Authorized by | In registry | Gap | |---|---|---|---|---|---| | <role> | <ID> | yes or no | <originator, head of office, or none recorded> | yes or no | <what to fix> | Template (fill names in your own system only) Standard list: <product>, <owner>, <recipient names> Copy log: <document ID>, <recipient name>, <authorized by>, <registry entry> Not covered: <anything asked that the bulletin does not state> ``` ## Rules - **A finding with no quoted line behind it is `UNVERIFIED`. Never state a handling rule from memory as confirmed.** - Remove names and personal details first. Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. - The audit and the template hold roles, names and document IDs only. Never copy the content of the report into the answer. - Quote the bulletin's own words. Keep "must" and "may" as written. Shorten only by cutting. - A request from the recipient is no authorization. Only the originator or the head of the receiving or originating department or office can authorize a copy. - Do not judge the classification level of the report. If the user's document does not state it, write `NOT IN DOCUMENT`. - If the question asks about a form, a system or a sanction the bulletin does not describe, write `NOT IN SOURCE`. - Send any fix beyond the quoted rules to the originator or the head of office. ## Copy-paste prompt (works in any assistant) > Check my distribution list and copy log for <product> against the UN Secretary-General's bulletin "ST/SGB/2007/6" on information sensitivity, classification and handling, at https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf. Use only that bulletin and my document below. Quote the lines on standard distribution lists, the special registry and duplication. For each recipient, say whether they are on the standard list, who authorized the copy, and whether it is in the registry. Then give me a clean list and copy log with roles and document IDs only. If the bulletin does not open, say so and do not answer from memory. Do not repeat any report content. <paste list with names replaced by roles>
donor-template
Map a partnerships officer's own master results sheet to one donor's reporting fields. The results are written once and reused for each funder. Use when someone has to fill a donor report or assessment template with results they have already written.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: donor-template description: Map a partnerships officer's own master results sheet to one donor's reporting fields. The results are written once and reused for each funder. Use when someone has to fill a donor report or assessment template with results they have already written. --- # Donor Template You give it one master results sheet and one donor template. It gives back one field map. Each field shows the line in your document that it came from. If your document does not cover a field, the field is marked GAP. **Inputs** - `results`: your own master results sheet or report text. Required. - `template`: the donor's reporting fields or form, pasted or attached. Required. - `donor`: optional. The funder's name, used only as a label. ## Why this exists Partnerships officers rewrite the same results for each funder. Every donor has its own format and its own assessment. The Joint Inspection Unit found that donor assessments "overlap to a large degree in terms of the content covered. While valued, such" assessments bring heavy transaction costs. It also found that relying on many funders "subjects the United Nations system to various forms of external governance frameworks with various reporting requirements". If you write results once and map them to each template, you stop rewriting. Each result also stays worded the same way for every funder. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | Joint Inspection Unit report on donor-led assessments of United Nations system organizations | The outcome section on how much donor assessments overlap. The section on global governance and the leadership role, on outside reporting requirements | `https://www.unjiu.org/sites/www.unjiu.org/files/jiu_rep_2017_6_english__0.pdf` | | The user's own document | The results to report and the donor fields to fill | Provided by the user | If the report does not open, say so in one line. Leave out the quotes and do not fill them from memory or a search summary. If the user's document cannot be read, stop and ask for it again. ## Steps - Read the user's results. Split them into single statements. Label each one with the heading or row it sits under in the document. - Read the donor template. List each field in the donor's own words and order. Note any word limit or format the template states. - For each field, find the statements that answer it. Copy the user's wording. Shorten only by cutting. - If no statement answers a field, write GAP and say in a few words what the field asks for. - If a statement fits more than one field, use it in each one and mark it as reused. - List the statements that no field used. This shows the officer what this donor does not ask for. - Open the Joint Inspection Unit report only if the user asks why one master sheet helps. Quote it exactly. ## Answer layout ``` <donor> template · mapped from <document name> | Donor field | Text from your results | From | |---|---|---| | <field in the donor's words> | "<your words>" | <heading or row in your document> | | <field> | GAP: <what the field asks for> | none | Not used by this template: <statements left over> Check before sending: <word limits or formats the template states> ``` ## Rules - **Take results only from the user's document. Never add a result, figure, partner or place that the document does not state.** - Keep the user's wording. Never turn a planned result into an achieved one. - A field with no matching text is GAP. Never fill it with a guess or a generic phrase. - Keep the donor's field names and their order exactly. - If a source cannot be opened, say so. Do not fill it from memory or a search summary. - Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. ## Copy-paste prompt (works in any assistant) > I will paste my master results sheet and a donor's reporting template. Map my results to each donor field. Use the donor's own field names and keep them in the same order. Take results only from my document and copy my wording, shortening only by cutting. For each field, show the heading or row in my document it came from. If nothing in my document answers a field, write GAP and say what the field asks for. Do not invent results, figures or partners. At the end, list the results this template did not use. Leave out anything marked confidential. If you cannot read my document, say so and do not fill it from memory.
external-marking-check
Check a draft that reuses text from a partner document marked confidential or otherwise sensitive. Confirms the partner's marking is kept or matched with equal or stronger protection, and flags where partner consent is needed before the text is declassified or made public. Use when a team wants to reuse partner material in a donor update, report or web page.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: external-marking-check description: Check a draft that reuses text from a partner document marked confidential or otherwise sensitive. Confirms the partner's marking is kept or matched with equal or stronger protection, and flags where partner consent is needed before the text is declassified or made public. Use when a team wants to reuse partner material in a donor update, report or web page. --- # External Marking Check One draft in, one check out. Each finding points to a passage in the user's own document and to the bulletin clause that governs it. If the bulletin cannot be opened, the check says so and gives no ruling. **Inputs** - `draft`: the user's own document, e.g. a donor update. Mark partner passages with a short label. Do not paste the partner's confidential text. - `partner_marking`: the marking on the partner's original, copied as printed, e.g. "Confidential". - `audience`: who will see the draft. Internal, donors only, or public. - `consent`: optional. Any written reply from the partner about reuse. ## Why this exists Partner proposals arrive stamped confidential. Teams then lift text into donor updates and the stamp gets lost on the way. The UN rule on information from outside sources is short. Keep the partner's marking or use one that protects at least as much. Before declassifying, weigh what the partner expects and ask them first where that is appropriate. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | UN Secretary-General's bulletin on information sensitivity, classification and handling | The clause on external information with prior markings. The clause on declassifying information from an outside source | `https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf` | | The user's own draft | The text to check and the markings it carries | Supplied by the user | The two clauses, as given in the evidence for this skill: > "3.2 Where information from an external source contains prior sensitivity markings, it shall retain those markings or shall be assigned a classification that provides a degree of protectio n greater than or equal to that of the entity that furnished the information." > "4.4 When declassifying information received from an outside source, the Organization shall give due regard to expect ations of confidentiality of that outside source and, if appropriate, shall seek the prior consent of the outside source." If the bulletin does not open, write `SOURCE NOT OPENED`. Do not fill it from memory or a search summary. ## Steps - Open the bulletin. Find both clauses and check that the wording matches the quotes above. If it does not open, say so and stop before any ruling. - Read the user's draft. List every labelled partner passage. If it is unclear whether a passage comes from the partner, ask the user. - For each passage, note the partner's marking and the marking the draft carries. - Apply the marking clause. A passage passes only if the draft keeps the partner's marking or carries a classification that protects at least as much. A public draft carries no protection. Partner text marked confidential in a public draft fails. - If the team wants to lower the marking or publish, apply the declassification clause. Say what the partner is likely to expect, based only on its marking and any reply supplied. Say whether consent should be sought first. Quote any consent reply. - Give a fix for each failed passage. Remove it, hold it until the partner agrees, or replace it with text the team owns that holds nothing from the partner's document. ## Answer layout ``` Draft: <title of the user's document> Audience: <audience> Bulletin opened: yes, or SOURCE NOT OPENED | Passage | Partner marking | Draft marking | Result | Clause | |---|---|---|---|---| | <label> | <as printed> | <as in draft> | kept, raised or fails | marking clause | Consent: <seek first, given, or not needed and why> "<quoted words from the declassification clause>" Fix: <one line per failed passage> Not covered: <anything asked that the bulletin does not state> ``` ## Rules - **No ruling without the bulletin open. If it cannot be opened, write `SOURCE NOT OPENED` and do not fill from memory.** - Quote the bulletin's words. Keep "shall" as "shall". Never soften it to "may" or "should". - Take the partner's marking as printed. Do not guess what it means beyond what the bulletin says. - Whether consent is appropriate is the team's call. Flag it every time partner text moves to a wider audience, and lean toward asking. - Never call partner text cleared for public use unless the user supplies a consent reply or states why consent is not appropriate. - If the question goes beyond the two clauses, such as who may declassify or how long a marking lasts, write `NOT IN SOURCE`. - Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. ## Copy-paste prompt (works in any assistant) > Check my draft below for text taken from a partner document marked <marking as printed>. The draft is for <internal, donors only, or public>. Use only the UN Secretary-General's bulletin at https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf and my draft. Quote the clause on external information with prior sensitivity markings and the clause on declassifying information from an outside source. For each labelled partner passage, say whether my draft keeps the partner's marking or one that protects at least as much, and whether I should seek the partner's consent before this goes out. Suggest a fix for any passage that fails. If you cannot open the bulletin, say so and do not answer from memory. I have left out the partner's confidential text and put a label where each passage sits.
logframe-qa
Check a user's own results framework or logframe line by line and say whether each line labelled as an outcome reads as an outcome or as an output with a new label. Each verdict gives a short reason, and background on why this happens is quoted from a Joint Inspection Unit report. Use when someone asks whether the outcomes in their logframe are really outputs, or wants them reworded before a progress report.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: logframe-qa description: Check a user's own results framework or logframe line by line and say whether each line labelled as an outcome reads as an outcome or as an output with a new label. Each verdict gives a short reason, and background on why this happens is quoted from a Joint Inspection Unit report. Use when someone asks whether the outcomes in their logframe are really outputs, or wants them reworded before a progress report. --- # Logframe QA Your own results framework in, a verdict on every outcome line out. Each verdict rests on your document's words and on the control question below. **Inputs** - `document`: required. The user's own results framework or logframe, pasted or attached. Only the lines to check are needed. - `team`: optional. What the team delivers with its own resources, for example workshops or guidance notes. It helps decide which lines the team controls. ## Why this exists Results frameworks often call something an outcome when the team can deliver it alone. A workshop held is in the team's control. What participants do afterwards is not. A Joint Inspection Unit report on results-based management found that organizations focus on outputs they control and sometimes label them outcomes. It links this to how the accountability system works. The same report names the pressure to report outcomes every reporting cycle, even when outcomes take longer than a cycle to appear. An output with an outcome label makes a report claim a change that nobody has shown. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | Joint Inspection Unit report on results-based management | Summary of challenges, where outputs the organization controls get identified as outcomes. A later list of challenges, where staff face pressure to report on outcomes every reporting cycle | `https://www.unjiu.org/sites/www.unjiu.org/files/jiu_rep_2017_6_english__0.pdf` | | The user's own document | The lines to check, and any definitions of output and outcome it gives | Pasted or attached by the user | Find these two passages in the report by searching for their words. - "identifying those as outcomes, given the nature of the accountability system, as well as" - "Disconnect between the longer-term outcome requirements of results-based management and the reporting cycle, with pressure to report on outcomes no matter the situation" If the PDF does not open, say so. Do not fill it from memory or from a search summary. The verdicts can still be given from the user's document alone. ## Steps - Open the report PDF. Copy its title and document symbol from the cover. Find both passages and check they read as quoted. If the PDF does not open, record `REPORT NOT OPENED` and go on. - Read the user's document. If it defines output and outcome, copy those definitions and use them in place of the control question. Say that you did. - List every line the document labels as an outcome. Keep its exact words. If the document has no labels, ask the user which lines are meant as outcomes. - Put the control question to each line. Could the team finish all of its own work and this line still not be true? If yes, the line reads as an outcome. If no, the team controls it and it reads as an output. - If the document does not say who has to change, or what the team delivers, mark the line `UNCLEAR` and name the missing piece. - For each output labelled as an outcome, write a rewrite that names the change the output should lead to. Use an actor the document already names. If it names none, write `ACTOR NOT IN DOCUMENT`. Keep the original line in the framework as an output. - If the report opened, add the first passage as the reason this pattern is common. Add the second passage only when the user mentions pressure to report outcomes every cycle. ## Answer layout ``` <document name> · checked <today's date> Report: opened, "<title from cover>" (or REPORT NOT OPENED and why) Test used: the control question (or the document's own definitions, quoted) | Line as written | Labelled | Reads as | Reason | |---|---|---|---| | "<line>" | outcome | OUTPUT | The team delivers this with its own work. | | "<line>" | outcome | OUTCOME | It needs <actor> to change. | | ... | | | | Rewrites: <each OUTPUT line, then a rewrite that names the actor and the change> Why it happens: "<passage>" JIU report, Summary of challenges Not covered: <anything asked that the document and the report do not answer> ``` ## Rules - **A verdict rests on the user's document and the control question. Anything said about why the pattern happens is a quoted passage from the report, or it is left out.** - Quote each line as the user wrote it. Do not tidy its wording in the table. - Never add a target, baseline, indicator value or date that the document does not contain. - Keep `UNCLEAR` when the document does not settle it. Do not guess a verdict to make the table look complete. - If the report did not open, say so at the top and quote nothing from it. - Do not judge the project or the team. The check is about labels only. - Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. ## Copy-paste prompt (works in any assistant) > Check the results framework I paste below. For every line labelled as an outcome, tell me if it reads as an outcome or as an output. Use this question. Could my team finish all of its own work and this line still not be true? If yes, it is an outcome. If no, my team controls it and it is an output. If my document defines output and outcome, use those definitions instead. Quote each line as I wrote it and give your verdict with a short reason. Write UNCLEAR where my document does not say enough. For each output labelled as an outcome, suggest a rewrite that names who should change, using only actors named in my document. For background, use only the Joint Inspection Unit report at https://www.unjiu.org/sites/www.unjiu.org/files/jiu_rep_2017_6_english__0.pdf and quote it word for word. If you cannot open it, say so and do not fill from memory. Do not add targets or dates that are not in my document.
meeting-actions
Turn your own meeting minutes into an action table with a draft follow-up for each action and the staff role who signs it off. Use when an intern or new joiner has written the minutes and must draft follow-ups they cannot send or sign in UNDP's name.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: meeting-actions description: Turn your own meeting minutes into an action table with a draft follow-up for each action and the staff role who signs it off. Use when an intern or new joiner has written the minutes and must draft follow-ups they cannot send or sign in UNDP's name. --- # Meeting Actions Your minutes in, an action table out, with every follow-up marked as a draft and a staff role named to sign it off. **Inputs** - `minutes`: your own minutes or working notes, pasted or attached, with names and personal details removed. - `meeting`: optional. The group's name as you want it in the header, for example the Result Group. - `sign_off`: optional. The staff role who approves your messages, for example your supervisor. ## Why this exists An intern may write the minutes and draft the follow-up. A UNDP Rwanda internship description lists this work: "ONE UN Rwanda, writing minutes and following up with communication, etc." The same document sets a limit: "Interns are not staff members and may not represent UNDP in any official". You can draft the follow-up. A staff member has to send and sign it. A quick draft from a chat assistant reads as if you could send it yourself. This skill puts the sign-off next to every message so nobody misses it. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | UNDP Rwanda internship description | Responsibilities and Accountabilities section: the intern writes minutes and follows up on communication. Conditions section: interns are not staff and may not represent UNDP officially | `https://www.undp.org/sites/g/files/zskgke326/files/migration/rw/Internship_ME.pdf` | | Your minutes or notes | The actions the meeting agreed and who owns each one | The text you paste or attach | If the PDF does not open, say so at the top of the answer and write `SOURCE NOT OPENED` in place of each quote. Do not fill the quotes from memory or a search summary. Do not use any other page, policy or template. ## Steps - Open the internship PDF. Find the Responsibilities and Accountabilities section and the Conditions section. Copy the two quoted lines word for word. If the file does not open, mark both `SOURCE NOT OPENED` and carry on with the minutes alone. - Read the minutes. If you see a name or other personal detail, stop. Ask the user to replace it with a role and paste the minutes again. - List each action the meeting agreed. Keep the minutes' own words. Give each action its own row. - Write the owner as a role taken from the minutes. Copy a due date only if the minutes state it. If they do not, write `NOT IN MINUTES`. - Draft a short follow-up for each action that needs a message. Start it with DRAFT. Write it for a staff member to send and leave the signature for them. - Fill the sign-off column with the staff role who approves and sends the message. Use `sign_off` if the user gave it. If nothing says who, write `ASK SUPERVISOR`. - Put anything the minutes leave unclear under Open questions. Leave it open for the user to check. ## Answer layout ``` <meeting> · minutes from user · source read <today's date> Task: "ONE UN Rwanda, writing minutes and following up with communication, etc." Responsibilities and Accountabilities Limit: "Interns are not staff members and may not represent UNDP in any official" Conditions (write SOURCE NOT OPENED in place of a quote that could not be checked) | Action, in the minutes' words | Owner role | Due | Draft follow-up | Signs off | |---|---|---|---|---| | <action> | <role> | <as written, or NOT IN MINUTES> | DRAFT: <short message> | <staff role, or ASK SUPERVISOR> | Open questions: <what the minutes leave unclear> All messages above are drafts. None has been sent or signed. ``` ## Rules - **Every follow-up is a draft. The intern does not send or sign anything in UNDP's name. A staff member signs off each message.** - Remove names and personal details first. Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. - Use only the internship PDF and the user's minutes. Add nothing the minutes do not say. - Quote the PDF word for word. If it does not open, say so. Never quote it from memory or from a search summary. - Put roles in the owner and sign-off columns. The user adds names after the draft leaves the AI tool. - A draft does not promise money or agree to anything for UNDP. If an action needs that, flag it in Open questions for the staff member who signs off. - If the user asks you to send or sign a message, decline and point to the sign-off column. ## Copy-paste prompt (works in any assistant) > I am an intern and I wrote these meeting minutes. I removed names and personal details. Turn them into a table of actions in the minutes' own words. For each action give the owner as a role, the due date only if the minutes state it, a short follow-up message marked DRAFT, and the staff role who signs it off. I may not send or sign anything in UNDP's name. Open https://www.undp.org/sites/g/files/zskgke326/files/migration/rw/Internship_ME.pdf and quote the line "Interns are not staff members and may not represent UNDP in any official". If you cannot open that file, say so and do not quote it from memory. Use no other source. If something in the minutes is unclear, list it as an open question.
metadata-read
Read the official SDG metadata PDF for any indicator code and return six fields in a fixed order (definition, unit, numerator and denominator, method of computation, limitations, custodian), each quoted word for word with its section heading. Use when someone asks what an SDG indicator measures, what it is divided by, how it is computed or who the custodian is.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: metadata-read description: Read the official SDG metadata PDF for any indicator code and return six fields in a fixed order (definition, unit, numerator and denominator, method of computation, limitations, custodian), each quoted word for word with its section heading. Use when someone asks what an SDG indicator measures, what it is divided by, how it is computed or who the custodian is. --- # Metadata Read One indicator code in, six quoted fields out, always in the same order. Every field says which section of the metadata it came from, and the section heading is copied as the PDF writes it. **Input** - `code`: an indicator code from the global SDG indicator framework, e.g. `3.1.1` ## Why this exists The wrong denominator is the most common way a correct number gets misread. The maternal mortality ratio in 3.1.1 is per 100,000 live births, and the metadata itself warns that the maternal mortality rate, a different measure, divides by person-years lived by women of reproductive age. In a recorded test without this skill, Claude got most of the 3.1.1 wording right and still put the custodian under `1.a "Organisation"`. The PDF names it in 0.g. Its 3.1.1 limitations were about vital registration and sibling histories, while 4.b of the PDF is about the ratio and the rate. For 6.1.1 it gave `Denominator (4.c): The total population.` Section 4.c of the 6.1.1 PDF never says what is divided by what. ## Sources (use these and nothing else) | What | Where | |---|---| | Metadata PDF | `https://unstats.un.org/sdgs/metadata/files/Metadata-<CC-TT-II>.pdf`. Each part of the code becomes two digits, so 3.1.1 becomes `03-01-01` and 6.1.1 becomes `06-01-01` | | Index of metadata files, if the PDF link fails | `https://unstats.un.org/sdgs/metadata/` | Do not use a WHO, UNICEF or World Bank page, a textbook, a news story or a summary from memory, even when it agrees. ## Steps 1. Open the PDF. Copy the `Last updated` date from the page header and the template line (`Harmonized metadata template - format version 1.1` or the like). 2. Find each field under its section. In the harmonized template these are: | Field | Section | |---|---| | Definition | `2.a. Definition and concepts (STAT_CONC_DEF)` | | Unit | `2.b. Unit of measure (UNIT_MEASURE)` | | Numerator and denominator | No section of its own. Look in 4.c first, then 2.a | | Method of computation | `4.c. Method of computation (DATA_COMP)` | | Limitations | `4.b. Comment and limitations (REC_USE_LIM)` | | Custodian | `0.g. International organisations(s) responsible for global monitoring (SDG_CUSTODIAN_AGENCIES)` | 3. Copy each section heading exactly as the PDF prints it, spelling included (`organisations(s)` is the PDF's own spelling). If the PDF uses an older template with other numbers, copy its headings and say so. 4. Quote the first sentence of the definition, the whole unit line and the first sentence of the limitations. For the method, quote the first sentence in 4.c that says how the value is computed. If 4.c opens with a list of production steps, skip it and take the sentence with the rule, such as the 6.1.1 line that the service level is "taken as the minimum of these three indicators". Fix only spacing broken by text extraction (`t aken` becomes `taken`). 5. For numerator and denominator, quote the one sentence that says what is divided by what. If no sentence in the PDF names both, write `NOT IN SOURCE` and quote the unit line instead. Do not build a fraction from the definition. ## Answer layout ``` <code> · <indicator name from 0.c> metadata last updated <date>, <template line> Definition "<quote>" <section heading> Unit "<quote>" <section heading> Numerator and denominator "<quote>" or NOT IN SOURCE <section heading> Method "<quote>" <section heading> Limitations "<quote>" <section heading> Custodian "<quote>" <section heading> Source: <PDF URL> · read <today's date> ``` ## Rules - **Cannot open the PDF: say "I can't open the metadata now" and give the link. Do not fill any field from memory.** - The order of the six fields never changes, even when the question asks about one of them. - A field the PDF does not have is `NOT IN SOURCE`. Never fill it from another indicator, another agency or what the field usually says. - Never convert a unit. Per 100,000 live births stays per 100,000 live births. Do not restate it as a percent or per woman. - When the PDF names a similar measure that uses a different denominator, add one line quoting that sentence with its section. For 3.1.1 that is the maternal mortality rate in 4.b. - Never invent a number. Every number in the answer appears in the PDF. ## Copy-paste prompt (works in any assistant) > Open the official UN metadata PDF for SDG indicator <code> at unstats.un.org/sdgs/metadata (the file is Metadata-CC-TT-II.pdf, so 3.1.1 is Metadata-03-01-01.pdf). Give me, in this order: definition, unit of measure, numerator and denominator, method of computation, limitations, custodian agency. Quote each one word for word and copy the section heading it sits under. If the PDF never says what is divided by what, write NOT IN SOURCE for that line. If you can't open the PDF now, say so and give me the link. Do not answer from memory. ## Running it as code `engines/ai/indicator_card.py` in the claude-for-sdg repository downloads the same PDF, keeps its SHA-256 and cuts sections by their numbers. Copies of the 3.1.1 and 6.1.1 PDFs used for the recorded test are in `data/anchors/skills/metadata-read/`.
partner-report-indicator-check
Check the indicators and deliverables in an implementing partner's narrative report against the baseline and end line in the user's own programme document. Use when someone asks whether a partner report matches the programme document, or whether a target the partner says was met is backed by the indicator as written.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: partner-report-indicator-check description: Check the indicators and deliverables in an implementing partner's narrative report against the baseline and end line in the user's own programme document. Use when someone asks whether a partner report matches the programme document, or whether a target the partner says was met is backed by the indicator as written. --- # Partner Report Indicator Check You give it one programme document and one partner report. It gives back one checked line per indicator, and each line says where its evidence was found. If the report says a target was met but gives no value, the line says so. **Inputs** - `programme_document`: the user's own programme or project document, or the part of it that holds the results framework. It must give each indicator, its baseline, its end line and its source of verification. - `partner_report`: the implementing partner's narrative report, or the section of it that reports on results. - `indicator`: optional. Name one indicator to check. If this is left out, check every indicator in the programme document. ## Why this exists Partner reports often say a target was met without giving the value, the unit or the source. A new officer then has to decide whether that sentence fits the baseline and end line agreed in the programme document. The UNDP terms of reference for an M&E intern name this task directly. This skill turns it into a fixed side-by-side check that shows the evidence for each line. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | UNDP Rwanda, Monitoring and Evaluation internship terms of reference | Its section on responsibilities and accountabilities states the task: "Review the programme documents and annual work plans, ensuring the indicators and deliverables of narrative reports from the implementing partners are in line with the baseline/end line information defined in the documents." It also names "the analysis of the collected data and the final reporting against project indicators." | `https://www.undp.org/sites/g/files/zskgke326/files/migration/rw/Internship_ME.pdf` | | The user's programme document | Indicators, baselines, end lines and sources of verification | Pasted or attached by the user | | The user's partner narrative report | What the partner says was achieved | Pasted or attached by the user | The terms of reference only describe the task. Every value in the answer comes from the user's two documents. If the UNDP file does not open, say so in one line and carry on with the user's documents. Do not describe what it says from memory. If a user document is missing or cannot be read, stop and ask for it. ## Steps - **Read the programme document.** List each indicator with its exact wording, baseline, end line, unit and source of verification. Note the page or section where each one sits. - **Read the partner report.** For each indicator, find the passage that reports on it. Copy the words and note the page or section. - **Match the indicator.** Check that the report talks about the same indicator. It must measure the same thing, in the same unit, for the same group of people or places. If the wording differs, quote both versions. - **Compare the value.** Put the reported value next to the baseline and the end line. Compare them only when both use the same unit. If the report gives no value, write `NO VALUE GIVEN`. - **Check the source.** Say whether the report names the source of verification that the programme document asks for. - **Check deliverables.** If the programme document or annual work plan lists deliverables, check each one the same way. - **Give a status** for each line. Use only these labels: `MATCHES`, `DIFFERS`, `NO VALUE GIVEN`, `NOT REPORTED`, `NOT IN PROGRAMME DOCUMENT`. - **List questions for the partner.** Write one line per gap, in plain words the officer can send. ## Answer layout ``` Indicator check · <programme name from the document> · <partner name from the report> UNDP terms of reference: opened / could not be opened | Indicator (programme document words) | Baseline | End line | Partner report says | Source named | Status | Where | |---|---|---|---|---|---|---| | <indicator> | <baseline> | <end line> | "<quote>" | yes / no | <status> | programme document <section>, report <section> | Questions for the partner: - <one line per gap> Not covered: <anything asked about that neither document states> ``` ## Rules - **"Target met" is a claim. It counts only when the report gives a value in the same unit as the end line. If there is no such value, the status is `NO VALUE GIVEN`.** - Quote both documents. Do not reword an indicator into a different one. - Never fill a missing baseline, end line or value from memory, from a similar project or from a guess. Write `NOT IN PROGRAMME DOCUMENT` or `NOT REPORTED`. - Do not work out new totals unless both documents give the parts in the same unit. If you do, show the working. - Mark a changed unit, group or place as `DIFFERS`, even when the partner calls the target met. - Do not judge the partner. Report what the documents say and what is missing. The supervisor decides. - If a source cannot be opened, say so. Do not fill it from memory or a search summary. - Remove names and personal details first. Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. ## Copy-paste prompt (works in any assistant) > I will paste two texts. The first is the results framework from my programme document. The second is an implementing partner's narrative report. For each indicator in the programme document, quote its wording, baseline, end line and source of verification. Then quote what the partner report says about it, and say where in the report it appears. Mark each line MATCHES, DIFFERS, NO VALUE GIVEN, NOT REPORTED or NOT IN PROGRAMME DOCUMENT. A statement that a target was met counts only if the report gives a value in the same unit as the end line. Do not fill any gap from memory and do not guess. End with a short list of questions I can send to the partner. If you cannot read one of the texts, tell me and stop.
procedure-step-check
Run a procedure-first check on the user's own anonymized timeline of a contested staff decision. Each step is checked against the procedural rules the user supplies, before any merits review. Use when an appeals or legal desk is preparing a case that may reach the ILO Administrative Tribunal and wants to find missed steps early.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: procedure-step-check description: Run a procedure-first check on the user's own anonymized timeline of a contested staff decision. Each step is checked against the procedural rules the user supplies, before any merits review. Use when an appeals or legal desk is preparing a case that may reach the ILO Administrative Tribunal and wants to find missed steps early. --- # Procedure Step Check One anonymized timeline in, one procedure check out. Every required step is matched to a timeline entry or marked missing. The merits come after. **Inputs** - `timeline`: the user's anonymized timeline of the contested decision, one event per line. - `rules`: the procedural text that applied, pasted from the user's own document. Each step keeps its reference as written there. - `decision`: optional. One line naming the contested decision. ## Why this exists Legal desks tend to start with the merits of a contested decision. The Joint Inspection Unit review of UNESCO appeals points to procedure: "Adverse judgments were largely due to procedural noncompliance and legal errors, underscoring the need for strengthened compliance and early dispute resolution." A skipped step is easier to find and fix before a case is argued than after. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | JIU review highlights | The section "Appeals processes and the ILO Administrative Tribunal", with the finding on adverse judgments | `https://www.unjiu.org/sites/www.unjiu.org/files/jiu-rep-2026-1_review_highlights.pdf` | | The user's own document | The anonymized timeline and the procedural rules text that applied | Supplied by the user | If the JIU file does not open, say so in the answer and write SOURCE NOT OPENED. Do not quote it from memory or from a search summary. The check still runs on the user's document. If the user gives no rules text, do not supply rules from memory. Write RULES NOT PROVIDED and list only the timeline events. ## Steps Run these in order. - **Open the JIU file.** Copy the quoted line and the section heading it sits under. If the file does not open, record that. - **Screen the user's document.** If it contains names or personal details, stop and ask the user to remove them. - **List the required steps.** Take each procedural step from the rules text, in its own words. Keep "shall" as required and "should" as expected. Note a time limit only as the rules state it. - **Match each step to the timeline.** Find the entry that shows the step was done. Quote that entry. - **Mark a status.** Done when an entry shows it. Missing when no entry shows it. Unclear when the entry is vague. Late when the rules set a time limit and the timeline shows it was passed. - **List the gaps.** Put Missing, Late and Unclear steps at the top. Name the record that would close each one. - **Only then turn to the merits.** Say the procedure check is complete and that merits review can start. Do not review the merits inside this skill. ## Answer layout ``` Procedure check · sources read <today's date> Finding: "<quoted JIU line>" JIU review highlights, <section heading> (or SOURCE NOT OPENED) | Step | Rule says | Timeline shows | Status | Rule reference | |---|---|---|---|---| | <step> | "<rule words>" | "<entry>" or none | Done, Missing, Unclear or Late | <reference in the user's document> | Gaps before merits: <Missing, Late and Unclear steps, with the record that would close each> Not covered: <anything asked that the opened documents do not state> ``` ## Rules - **Remove names and personal details first. Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows.** - Every required step comes from the user's rules text. Never add a step from memory or from general practice. - Quote the rules and the timeline. Shorten only by cutting. Never turn a "should" into a "shall". - A step with no timeline entry is Missing. Do not assume it happened. - Do not predict how the Tribunal would rule. Do not give a view on the merits. - If the question needs a rule, a time limit or a form that the user's document does not hold, write NOT IN SOURCE. - Never invent a number, a time limit or a date. ## Copy-paste prompt (works in any assistant) > Check the procedure in this anonymized case before any merits review. Here is my timeline and the procedural rules that applied: <paste>. Use only my text and the JIU review highlights at https://www.unjiu.org/sites/www.unjiu.org/files/jiu-rep-2026-1_review_highlights.pdf. List each step the rules require, in their own words. For each step, quote the timeline entry that shows it was done, or mark it Missing, Late or Unclear. Put the gaps first. If you cannot open the JIU file, say so and do not quote it from memory. Do not add rules I did not give you. Do not give a view on the merits or the likely outcome.
question-batch-tracker
Turn a batch of reviewer questions in the user's own document into a tracker that logs each question with its batch, owner, due date and response status. Use when a Party's focal point gets another batch of reviewer questions and has to restart coordination with national agencies.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: question-batch-tracker description: Turn a batch of reviewer questions in the user's own document into a tracker that logs each question with its batch, owner, due date and response status. Use when a Party's focal point gets another batch of reviewer questions and has to restart coordination with national agencies. --- # Question Batch Tracker One document in, one tracker out. Each question gets a batch, an owner, a due date and a status. Anything the document does not state is marked OPEN. **Inputs** - `document`: required. The user's own file with the reviewer questions. It can be pasted text, an export from the question and answer platform, or an earlier version of this tracker. - `batch`: optional. The name of the new batch, as the reviewers labelled it. - `owners`: optional. The agencies or colleagues who usually answer each topic. ## Why this exists Reviewers send questions in several batches. Each batch starts the chase across agencies again, and there is little time to consult. The online platform does not help much. A UNFCCC document records both problems as Parties reported them. A plain tracker kept outside the platform shows who owes which answer and by when. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | UNFCCC CMA document (PDF) | A table of challenges Parties reported. It names the batch problem and the platform problem in the two lines quoted below | `https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf` | | The user's own document | The questions, any owners, any due dates and any answers already drafted | Supplied by the user | The two lines to quote: - "Multiple batches of questions requiring repeated internal coordination and limited time for inter-agency consultation and preparation of responses" - "Technical issues with online question and answer platforms (e.g. character limits and attachment size restrictions, limited functionality for notifications, follow-up exchanges and tracking responses)" The UN document server sometimes serves a check page instead of the PDF. If the file does not open, say so. Do not fill the quotes from memory or a search summary. ## Steps - Open the UNFCCC PDF and find the two lines above. If it does not open, write `SOURCE NOT OPENED` in the header and work from the user's document only. - Read the user's document. Copy each question in the reviewers' own words. Shorten only by cutting. - Keep any question ID the reviewers gave. If there is none, use the batch name and a letter. - Take the owner from the document or from `owners`. If there is none, write `OPEN: no owner`. - Copy a due date only if the document states one. If there is none, write `OPEN: no due date`. - Set the status to one of these: Not started, With owner, Draft in, Cleared, Sent. Take it from the document. If the document is unclear, use Not started. - If the document holds an earlier tracker, keep the earlier batches and add the new batch below them. Flag any older question that is still not sent. - If a draft answer is long or needs an attachment, flag it `CHECK PLATFORM LIMITS`. Use the limits the platform itself shows. - End with a chase list. Give one line per owner with the question IDs they still owe. ## Answer layout ``` <batch> · tracker built <today's date> · source: opened or SOURCE NOT OPENED Why: "<quoted line on batches>" | ID | Question (reviewers' words) | Owner | Due | Status | Flag | |---|---|---|---|---|---| | <id> | <question> | <owner or OPEN> | <due or OPEN> | <status> | <flag or blank> | Chase: <owner>: <IDs still owed> Open items: <questions with no owner or no due date> ``` Keep the answer short. Use one row per question. ## Rules - **An owner or due date the document does not state is `OPEN`. Never fill one from memory or a guess.** - Keep the reviewers' wording for each question. Do not merge questions. - Quote the UNFCCC lines word for word, or leave them out. - Do not state the platform's character or file size limits. The source does not give them. - Do not draft the technical answers unless the user asks. This skill tracks them. - Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. ## Copy-paste prompt (works in any assistant) > I am a focal point and I just received another batch of reviewer questions. From the document I paste below, build a tracker table with one row per question. Give the reviewers' own wording, the owner, the due date, the status (Not started, With owner, Draft in, Cleared or Sent) and a flag. Take owners and due dates only from my document. Where one is missing, write OPEN and do not guess. Keep any earlier batches I include and flag questions that are still not sent. Mark answers that are long or need attachments with CHECK PLATFORM LIMITS. End with a chase list for each owner. If you cite the UNFCCC document at https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf, quote it word for word. If you cannot open it, say so. Do not answer from memory.
redd-annex-presubmission-check
Check a draft REDD+ technical annex to a Biennial Transparency Report (BTR) before it leaves the ministry. The check finds gaps, open placeholders, broken cross-references and mismatches in the user's own document. It also flags timing risk, using the UNFCCC secretariat's finding on annexes that arrive late. Use when a national team or a transparency officer asks whether an annex is ready to send.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: redd-annex-presubmission-check description: Check a draft REDD+ technical annex to a Biennial Transparency Report (BTR) before it leaves the ministry. The check finds gaps, open placeholders, broken cross-references and mismatches in the user's own document. It also flags timing risk, using the UNFCCC secretariat's finding on annexes that arrive late. Use when a national team or a transparency officer asks whether an annex is ready to send. --- # REDD+ Annex Pre-submission Check One draft annex goes in and one ready-or-not list comes out. Every flag points to a place in the user's document or to a quoted line from the UNFCCC report. **Inputs** - `annex`: required. The user's own draft technical annex, pasted or attached. - `send_plan`: optional. The user's own words on when the team plans to send the annex and when review week falls. - `btr_text`: optional. The main BTR text that goes with the annex. The check uses it to confirm that the two agree. ## Why this exists The secretariat has reported that some annexes arrived just before review week. The LULUCF experts then had little time to prepare. The secretariat also expects more annexes in the second BTR cycle. An annex that is complete and easy to follow when it arrives gives the experts the time they need. A check before the file leaves the ministry catches gaps that would otherwise come up during review. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | UNFCCC secretariat report on the technical analysis | The finding that some annexes came shortly before review week. The expectation of more annexes in the second BTR cycle. | `https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf` | | The user's own draft annex | Everything the check says about content, structure and gaps | Supplied by the user | | The user's BTR text, if given | Values and terms the annex must match | Supplied by the user | The UN document link sometimes returns an error page or a login page instead of the PDF. If it does not open, say so in the answer. Do not quote it from memory or from a search summary. Run the document checks anyway and mark the timing line `SOURCE NOT OPENED`. ## Steps - Open the UN document. Find this line and copy it exactly with its page: "In some cases, technical annexes were submitted shortly before the corresponding review weeks, leaving limited time for the secretariat and LULUCF experts to prepare for the technical analysis." - Find this line too and copy it with its page: "The number of technical annexes submitted is likely to increase in the second cycle of BTRs as developing country Parties seek to demonstrate their REDD+ results" - Read the whole annex before writing anything. - List every section heading in the annex. If it has a table of contents, check that each listed section is there. - Find every placeholder. Look for TBD, to be added, XX, empty tables, highlighted notes and comments left in the text. - Find every cross-reference. Check that each table, figure, annex or section it points to exists in the document. - Flag any passage that sends the reader to another document without naming that document. - If `btr_text` is given, compare each value and term the annex shares with it. List each mismatch with both locations. - If `send_plan` is given, compare it with the quoted finding. If the plan puts the annex close to review week, flag the timing risk and quote the line. Do not state a deadline. The opened sources do not set one. - Sort the flags. Blockers come first, then fixes, then notes. ## Answer layout ``` Annex check · <annex title as written in the document> Source opened: yes, or SOURCE NOT OPENED and why Timing: <flag, or no plan given> "<quoted line>" UN document, page <n> | Flag | Where in the annex | What is wrong | Fix | |---|---|---|---| | Blocker | <section or table> | <placeholder, broken reference or mismatch, quoted> | <one short action> | | Fix | ... | ... | ... | | Note | ... | ... | ... | Ready to send: yes, or no with the count of blockers Not checked: <anything asked that the opened sources do not cover> ``` ## Rules - **If the UN document does not open, say so. Never fill the quote or the page from memory.** - Judge the annex only against itself, the BTR text the user gives and the quoted finding. - The opened sources do not list what a REDD+ annex must contain. If the user asks about required content, write `NOT IN SOURCE`. - Never invent a deadline, a review week date, a required section or a number of tables. - Quote the user's document exactly when you cite a problem. - Do not change technical values in the annex. Point to them and say what does not match. - Keep the user's document inside this conversation. Do not send it to any other service or tool. - Write in plain English with short sentences. ## Copy-paste prompt (works in any assistant) > Check my draft REDD+ technical annex before we send it with our Biennial Transparency Report. Use only my annex, the BTR text I give you and the UNFCCC document at https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf. List every placeholder, every cross-reference that points to something missing, every passage that sends the reader to an unnamed document and every value that differs from my BTR text. Quote my text for each problem and say where it is. Sort the list into blockers, fixes and notes. If my send plan puts the annex close to review week, quote the UN document's line about annexes that arrived shortly before review weeks, with its page. If you cannot open the UN document, say so and do not quote it from memory. Do not invent deadlines or required sections.
ref-year-check
Check a country comparison table (CSV or pasted rows) for one SDG indicator against the UN SDG Global Database, and flag every row whose value comes from a different reference year than the rest, or from a partial row such as urban only. Shows the database year of each value. Use before a table with "latest" values goes into a brief, a slide or a donor report.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: ref-year-check description: Check a country comparison table (CSV or pasted rows) for one SDG indicator against the UN SDG Global Database, and flag every row whose value comes from a different reference year than the rest, or from a partial row such as urban only. Shows the database year of each value. Use before a table with "latest" values goes into a brief, a slide or a donor report. --- # Reference Year Check One table in, the same table back with a database year next to every value. A table that compares countries is only fair when every value is from the same year and the same population. **Inputs** - `table`: CSV or pasted rows with a country column and a value column. A year column is optional. - `code`: the SDG indicator code the values claim to be, e.g. `6.1.1`. If the table does not say, ask. ## Why this exists Comparison tables are built over weeks from several downloads. Each value was "latest" on the day it was copied, so one table can hold several years and still carry one year in its header. The database also holds partial rows. For some countries the only rows for a series are urban or rural, and a value copied from them reads like a national figure. ## Sources (use these and nothing else) | What | Where | |---|---| | Series codes for the indicator | `https://unstats.un.org/sdgapi/v1/sdg/Indicator/List`, the entry for the exact code | | Country codes | `https://unstats.un.org/sdgapi/v1/sdg/GeoArea/List` | | Values with years | `https://unstats.un.org/sdgapi/v1/sdg/Series/Data?seriesCode=<SERIES>&areaCode=<M49>&pageSize=1000` | Do not use an agency portal, a report table or memory to find a year. The years in this check come from `timePeriodStart` in the database rows. ## Steps 1. Find the series for the code. If there are several, ask which one the table uses, or check each and say so. 2. Get the M49 code for every country in the table, using the names as the database writes them. 3. Pull every row for each country. Keep the total rows (`ALLAREA`, `BOTHSEX`, `ALLAGE` where those dimensions exist). Note when a country has no total row. 4. For each table value, find the database year whose value matches it. Round the database value to the table's number of decimals before comparing, and never round the table value. 5. Take the year most rows share as the table's reference year. Flag every row that differs from it, every row that matches only a partial row, every row with no match and every empty cell. 6. For each flagged row, also give the database value for the reference year, if there is one, so the author can replace it. ## Answer layout ``` <code> · series <SERIES> · SDG Global Database, read <today's date> Reference year: <year> (<n> of <total> rows) | Country | Table value | Table year | Database year of that value | Row used | Database value for <ref year> | Flag | |---|---|---|---|---|---|---| | <country> | <v> | <y> | <year> | ALLAREA | <value> (<nature>) | OK | | <country> | <v> | <y> | <year> | ALLAREA | <value> (<nature>) | YEAR DIFFERS | | <country> | <v> | <y> | <year> | URBAN only | NOT IN SOURCE for national | PARTIAL ROW | | <country> | <v> | <y> | no match | ALLAREA | <value> (<nature>) | NO MATCH | | <country> | empty | <y> | none | none | NOT IN SOURCE | NO DATA | <n> rows flagged. ``` ## Rules - **Cannot reach the API: say "I can't check the database now", give the link and flag nothing. Do not guess years.** - A row with no total in the database is `NOT IN SOURCE` for the national value. Never use an urban or rural value in its place, and never average urban and rural. - A value that matches no year is `NO MATCH`. Do not pick the nearest year. - If two years round to the same table value, list both years and flag the row. - Keep the `Nature` attribute (estimated, modelled, country data) next to every database value. - Never fill an empty cell. Report what the database has for it, or `NOT IN SOURCE`. ## Copy-paste prompt (works in any assistant) > Here is a country comparison table for SDG indicator <code>. For every row, look up the country in the UN SDG Global Database (unstats.un.org/sdgapi, Series/Data) and tell me which year the value in my table belongs to. Use only the national total row. Tell me if a country has only urban or rural rows. Give the year most rows share, and flag every row from another year, every value that matches no year and every empty cell. For each flagged row, give the database value for the shared year. Do not guess any year from memory. If you can't open the database now, say so. ## Demo `demo/water_6.1.1.csv` is an eight-country 6.1.1 table where every row claims 2024. `demo/answer_key.md` lists what is planted in it and what comes from the database itself. Copies of the API responses with their SHA-256 are in `data/anchors/skills/ref-year-check/` in the claude-for-sdg repository.
review-finding-drafter
Turn a technical expert reviewer's own working notes on an issue into a side-by-side card that shows the same issue drafted as a question, as a recommendation and as an encouragement. Use when a first-time reviewer under the Enhanced Transparency Framework has a blank page and needs a worked example built from their own document.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: review-finding-drafter description: Turn a technical expert reviewer's own working notes on an issue into a side-by-side card that shows the same issue drafted as a question, as a recommendation and as an encouragement. Use when a first-time reviewer under the Enhanced Transparency Framework has a blank page and needs a worked example built from their own document. --- # Review Finding Drafter You give it your notes and get back a card. The card drafts the same issue as a question, a recommendation and an encouragement, side by side. Every fact in a draft comes from your own document. **Inputs** - `document`: required. Your own working notes or draft review text. Paste it or attach it. - `issue`: optional. The point in the document you want drafted. If it is missing, take the first issue the notes raise and say which one you took. ## Why this exists First review. Blank page. A UNFCCC report on the technical expert review process names this gap. It lists "Limited practical examples and guidance for drafting questions, areas of improvement, recommendations, encouragements and capacity-building needs, particularly for first-time technical expert reviewers". The same report gives lead reviewers the job of keeping each review consistent. A draft card gives the new reviewer a worked example from their own case to take to the lead reviewer. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | UNFCCC report on the technical expert review process, the table of challenges | The line on limited examples and guidance for first-time technical expert reviewers | `https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf` | | Same report, the passage on lead reviewers | "The LRs ensure consistency and quality in the TER process, facilitate communication between the Parties and the TERT, and provide guidance to the secretariat in compiling and editing the final TERRs." | Same file | | The user's own document | The facts of the issue. This is the only source for the content of the drafts | Supplied by the user | If the UNFCCC file does not open, say so at the top of the answer. Do not restate its content from memory or from a search summary. The card can still be drafted from the user's document. ## Steps - Open the UNFCCC file. Find the two passages above and copy them word for word. If a passage is not in the opened file, write `NOT FOUND` next to it. - Read the user's document. Pick the issue. List only the facts the document states about it. Write any missing fact as a gap. - Draft the question. It asks the Party for the missing information. Keep it short and specific. Name the gap in the document's own words. - Draft the recommendation. It states what the Party should do to fix the gap in its next report. Use only facts from the document. - Draft the encouragement. It invites the Party to make an improvement it may choose to make. Keep it softer than the recommendation. - Under each draft, write a short line on why that label fits. Mark these lines as drafting guidance, since the quoted passages do not define the labels. - End with a line telling the reviewer to check the wording with the lead reviewer, and quote the lead reviewer passage. ## Answer layout ``` Issue: <the issue in the document's words> Source: opened / UNFCCC source not opened | Question | Recommendation | Encouragement | |---|---|---| | <draft> | <draft> | <draft> | | Why: asks for missing information | Why: names a fix the Party should make | Why: invites an optional improvement | Gaps in your notes: <facts the drafts need that the document does not state> Check with your lead reviewer: "<quoted lead reviewer passage>" Label guidance is drafting help. It is not a UNFCCC definition. ``` ## Rules - **Use only facts in the user's document. Never add a method, gas, sector, figure or name the document does not state.** - Quote the UNFCCC passages word for word. Never present a paraphrase as a quote. - The quoted passages list the drafting types without defining them. Say so on the card. Do not claim an official definition from memory. - If the UNFCCC file cannot be opened, write `UNFCCC source not opened` and do not fill its content from memory or a search summary. - Keep every draft neutral and polite. Address the report and the information. Do not blame the Party. - Use plain English and short sentences. - Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows. ## Copy-paste prompt (works in any assistant) > I am a first-time technical expert reviewer. Using only my notes below, draft the same issue three ways, side by side: as a question to the Party, as a recommendation and as an encouragement. Under each draft, add a short line on why that label fits, and say that this is drafting help and not an official definition. Do not add any fact my notes do not state. List any gaps in my notes. If you can open the UNFCCC report at https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf, quote its passage on lead reviewers word for word and remind me to check my wording with my lead reviewer. If you cannot open it, say so and do not fill it from memory. My notes: <paste your notes>
roster-availability-tracker
Turn the user's own expert roster into a sheet that flags when each expert last confirmed availability, who still needs a fresh check, and who the backups are. Every rule about timing is traced to a quoted line in the secretariat report on technical expert reviews. Use when a secretariat officer is building a technical expert review team and experts are declining because their work changed after the availability survey.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: roster-availability-tracker description: Turn the user's own expert roster into a sheet that flags when each expert last confirmed availability, who still needs a fresh check, and who the backups are. Every rule about timing is traced to a quoted line in the secretariat report on technical expert reviews. Use when a secretariat officer is building a technical expert review team and experts are declining because their work changed after the availability survey. --- # Roster Availability Tracker Your roster in, one checked sheet out. Each expert gets a status, a backup and a next action, and every timing rule is backed by a quote. **Inputs** - `roster`: required. The user's own document. A sheet or table with each expert, their area, when they last confirmed availability, and any named backup. - `review`: optional. Which review the team is for and its review week, if the user knows it. ## Why this exists An expert who said yes in the availability survey may not be free when the invitation arrives. The secretariat report says the most common reason for declining was "changes in work commitments since responding to the availability survey (50.6 per cent)". The team also has a fixed lead time. A roster that only shows survey answers hides who needs a fresh check and who has no backup. ## Sources (use these and nothing else) | Document | What it holds | File | |---|---|---| | Secretariat report on technical expert reviews | Section C on team composition: when the team must be formed. The reasons experts gave for declining. The section on challenges: limited availability and high rejection rates | `https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf` | | The user's own roster document | Experts, areas, last confirmation, backups | Supplied by the user | If the report does not open, say so in the first line of the answer. Do not fill the timing rule or the decline reasons from memory or a search summary. Still build the sheet from the roster and mark each timing line `SOURCE NOT OPENED`. ## Steps - Open the report. Find and copy these lines word for word: "The secretariat forms the TERT for each TER at least 10 weeks prior to the review" and the line on "changes in work commitments since responding to the availability survey". - Read the user's roster. Keep each expert's label exactly as the user wrote it. - Give each expert one status. `Reconfirmed` if they confirmed after the survey. `Survey only` if their last confirmation was the survey. `No record` if the roster does not say. - Mark every `Survey only` and `No record` expert as needing a fresh check before the team is formed. - For each expert, name the backup from the roster. If there is none, write `NO BACKUP`. - If the user gave a review week, state the latest day to form the team using the quoted lead time. If not, give the rule as quoted and ask for the review week. - List any expert who appears as a backup for more than one person. One person cannot cover two gaps at once. ## Answer layout ``` Roster check · report opened: yes or SOURCE NOT OPENED Team deadline: "<quoted line>" <latest day, or ask for review week> | Expert | Area | Last confirmed | Status | Backup | Next action | |---|---|---|---|---|---| | <label> | <area> | <as in roster> | Survey only | <name or NO BACKUP> | Reconfirm now | Check first: <experts marked Survey only or No record> Gaps: <experts with NO BACKUP, and backups used twice> Why reconfirm: "<quoted decline reason>" ``` ## Rules - **Remove names and personal details first. Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows.** - Use only the report and the user's roster. Do not add experts, backups or confirmations the roster does not state. - Quote the report for every timing rule. If the report did not open, write `SOURCE NOT OPENED` and do not supply the rule from memory. - Treat a survey answer as a starting point. An expert is `Reconfirmed` only when the roster shows a check after the survey. - Keep the user's own labels and wording for each expert and area. - If the user asks about funding, eligibility or invitation rules the report does not state, write `NOT IN SOURCE`. - Keep sentences short. Put the next action for each expert in plain words. ## Copy-paste prompt (works in any assistant) > Here is my expert roster for a technical expert review team. Remove any names and personal details before you read it. For each expert, tell me when they last confirmed availability, whether that was only the availability survey, who their backup is, and what I should do next. Flag anyone with no backup and any backup listed for more than one person. Use only my roster and the secretariat report at https://documents.un.org/api/symbol/access?s=FCCC/PA/CMA/2026/10&l=en&t=pdf. Quote the report's line on when the team must be formed and its line on why experts decline. If you cannot open the report, say so and do not answer those points from memory.
indicator-card
Build a one-page, citable card for any SDG indicator and country. Use when someone gives an SDG indicator code (like 6.1.1) and a country and wants the definition, the latest official value, the custodian agency and the methodological catch, each line traced to its official source.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: indicator-card description: Build a one-page, citable card for any SDG indicator and country. Use when someone gives an SDG indicator code (like 6.1.1) and a country and wants the definition, the latest official value, the custodian agency and the methodological catch, each line traced to its official source. --- # SDG Indicator Card Two inputs, one card. Every line on the card says where it came from. **Inputs** - `code`: an indicator code from the global SDG indicator framework, e.g. `6.1.1` - `country`: a country name as the SDG Global Database writes it, e.g. `Ghana` ## Why this exists An assistant asked for an indicator value from memory gives a number, a rough year and a hedge. In a recorded test without this skill, Claude said Ghana's 6.1.1 figure was "about 38%" for "around 2022". The SDG Global Database read on 20 October 2026 has 42.9 percent for 2024, estimated. For Kenya it said "somewhere around 30–40%". The database has no Kenya row for the series `SH_H2O_SAFE` at all. ## Sources (use these and nothing else) | What | Where | Cite as | |---|---|---| | Definition, custodian, method | `https://unstats.un.org/sdgs/metadata/files/Metadata-<CC-TT-II>.pdf` (6.1.1 becomes `06-01-01`) | `metadata <code> §<section>, last updated <date>` | | Series list for the code | `https://unstats.un.org/sdgapi/v1/sdg/Indicator/List` | `SDG API, Indicator/List` | | Country code | `https://unstats.un.org/sdgapi/v1/sdg/GeoArea/List` | `SDG API, GeoArea/List` | | Values | `https://unstats.un.org/sdgapi/v1/sdg/Series/Data?seriesCode=<SERIES>&areaCode=<M49>&pageSize=1000` | `SDG Global Database, <SERIES>, <year>` | ## Steps 1. Open the metadata PDF. Copy the **last updated** date from the page header. 2. Copy the custodian agencies from **§0.g** and the first sentence of the definition from **§2.a**, word for word. 3. Read **§4.c** (computation), **§4.f** (missing values) and **§4.g** (aggregation). Pick the one rule a reader is most likely to get wrong. Quote it word for word with its §. 4. Get the series codes for the indicator, then the country's rows for each series. 5. Use the total row (`ALLAREA`, `BOTHSEX`, `ALLAGE` where those dimensions exist). Report the latest year, the value, the unit and the `Nature` attribute (estimated, modelled, country data). ## Card layout ``` <code> · <official indicator name> <country> · <value><unit> (<year>, <nature>) SDG Global Database, <SERIES> Definition "<first sentence of §2.a>" metadata §2.a, <last updated> Custodian <agencies> metadata §0.g The catch "<quoted rule>" metadata §4.x Source <source field of the data row> ``` ## Rules - **No value in the database: write `NOT IN SOURCE` and say which series was empty.** Never fill it from a news story, a report or a guess. - Never average, sum or convert values. If the method says how a figure is built, quote that sentence instead. - If an indicator has several series, list each series code. Do not choose one silently. - Take series only from the entry for the exact code in `Indicator/List`. A series filed under a neighbouring code, such as `SH_SAN_HNDWSH` under 6.2.1, is not a series of 6.1.1. - Quotes are word for word. If the PDF text is broken by extraction (`t aken`), fix only the spacing. - Keep the `Nature` label. An estimated or modelled value is not the same as country-reported data. ## Copy-paste prompt (works in any assistant) > For SDG indicator <code> in <country>, take the value only from the UN SDG Global Database (unstats.un.org/sdgapi) and the definition and custodian only from the official metadata PDF (unstats.un.org/sdgs/metadata). Quote the definition word for word with its section number. Give the year, the unit and whether the value is estimated, modelled or country data. If the database has no row for this country, write NOT IN SOURCE and name the empty series. If you can't open these pages now, say so and give me the links instead of a number. Do not answer from memory. ## Running it as code `engines/ai/indicator_card.py <code> <country>` in the claude-for-sdg repository does steps 1 to 5 and keeps a copy of every source it read, with its SHA-256.
sensitivity-precheck
Check a UN staff member's own draft before it goes into an AI tool. Help them decide whether it is sensitive, mark it, and keep classified material on protected channels, with every rule quoted from the Secretary-General's bulletin on information sensitivity. Use when someone is about to paste a note, email or draft into an AI tool and has not yet decided how it is classified.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
---
name: sensitivity-precheck
description: Check a UN staff member's own draft before it goes into an AI tool. Help them decide whether it is sensitive, mark it, and keep classified material on protected channels, with every rule quoted from the Secretary-General's bulletin on information sensitivity. Use when someone is about to paste a note, email or draft into an AI tool and has not yet decided how it is classified.
---
# Sensitivity Precheck
One draft in, one pre-paste check out. Every rule is quoted from the bulletin. The person who wrote or received the document makes the classification call.
**Inputs**
- `document`: the user's own draft, note or email, with names and personal details already removed.
- `marking`: optional. Any classification the document already carries.
- `origin`: optional. Whether the user wrote it or received it from an outside source.
## Why this exists
New staff often paste a draft into an AI tool to fix the wording before anyone has asked whether it is sensitive. The bulletin gives that decision to the person who created or received the information. It also limits how classified material may travel electronically. An AI tool cannot make this decision for the user. The user decides and marks first, and only then picks a channel.
## Sources (use these and nothing else)
| Document | What it holds | File |
|---|---|---|
| Secretary-General's bulletin on information sensitivity | The section on identification and markings (who decides and marks). The handling section and its line on electronic transmission. The section that lists the classification levels | `https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf` |
| The user's own document | The text to check | Supplied by the user |
The PDF usually opens with a plain download. If it does not open, say so in the first line of the answer. Then use only the two quotes in the Steps below and write `NOT OPENED` for everything else, including the list of levels. Do not fill the gaps from memory or a search summary.
## Steps
- Open the bulletin. Copy its symbol and date from the first page.
- Find the passage on identification and markings. Check that it reads: "Section 3 Identification and markings 3.1 The originator of the information concer ned, or its recipient if the information is received from an outside source, under the overall supervision and guidance of the head of department or office, shall decide whether the information is sensitive and mark it with the appropriate classifi cation as detailed in section 4 below."
- Find the line on electronic transmission. Check that it reads: "(f) Electronic transmission of classifi ed information shall be performed only through the use of protected means of communication, in accordance with section 5.4 below."
- Open the section the first quote points to ("as detailed in section 4 below"). List each classification level in the bulletin's own words, with its own definition. If that section does not open, write `NOT OPENED`.
- Read the user's document. If it already carries a classification marking, or the user says it is classified, stop. Quote the transmission line and do not work on the text.
- Otherwise, go through the draft and find each passage that may make it sensitive under the definitions you opened. Quote the passage. Give the reason in a few words, tied to the bulletin's wording. Mark each one `YOUR CALL`.
- Close with the marking step. The user decides the level and marks the draft before sharing it. Point them to their head of department or office for guidance, since the bulletin places the decision under that supervision.
## Answer layout
```
Sensitivity precheck · bulletin <symbol>, <date> · read <today's date>
Who decides: "<quoted line>" <section and paragraph>
How it may travel: "<quoted line>" <section and paragraph>
Levels: <each level in the bulletin's words> (or NOT OPENED)
| Passage in your draft | Why it may be sensitive | Bulletin wording | Status |
|---|---|---|---|
| "<quoted text>" | <a few words> | "<short quote>" | YOUR CALL |
Your next step: decide the level, mark the draft, then choose the channel.
Not covered: <anything asked that the opened bulletin does not state>
```
## Rules
- **Remove names and personal details first. Use your own working notes, leave out anything marked confidential, and use the AI tool your organization allows.**
- **Never decide the classification for the user. Flag and quote. The originator or recipient makes the call.**
- Never tell the user a draft is safe to paste. Say what the bulletin requires and leave the decision with them.
- If the document is marked or described as classified, stop and quote the transmission line.
- Quote the bulletin for every rule. Keep its "shall" as "shall".
- Use the bulletin's names for the levels. Never invent a level or a definition.
- If the user asks which AI tools count as protected means of communication, or which tool their organization allows, write `NOT IN SOURCE`. Treat no tool as protected unless an opened document says so.
- If the bulletin does not open, say so first. Never fill a section from memory.
- Write in plain English with short sentences.
## Copy-paste prompt (works in any assistant)
> Before I share my draft anywhere, help me check it for sensitivity. I have removed names and personal details. Use only the Secretary-General's bulletin at https://documents.un.org/api/symbol/access?s=ST/SGB/2007/6&l=en&t=pdf and my draft below. Quote the bulletin's lines on who decides and marks sensitivity, on electronic transmission of classified information, and its list of classification levels. Then quote each passage in my draft that may be sensitive, say why in a few words, and mark it as my call. Do not decide the level for me and do not tell me the draft is safe to paste. If you cannot open the bulletin, say so and do not answer from memory. My draft: <paste here>
latest-edition
Answer any request for a figure from a UN flagship report (HDI, SDG Report, World Population Prospects, SOFI, World Health Statistics and the like) from the latest edition only. Use when someone asks for a number "from the latest report", or names a report without an edition. Names the edition and its publication date, reads the publisher's current file, and refuses to give a number from memory.
Use it
- In Claude, add the file as a skill (Customize, then Skills).
- Check the source line under each value before you quote it.
Can't use Claude at work? Paste the rules section into the assistant you are allowed to use, and give it the source links.
Read the file
--- name: latest-edition description: Answer any request for a figure from a UN flagship report (HDI, SDG Report, World Population Prospects, SOFI, World Health Statistics and the like) from the latest edition only. Use when someone asks for a number "from the latest report", or names a report without an edition. Names the edition and its publication date, reads the publisher's current file, and refuses to give a number from memory. --- # UN Latest Edition One question in, one answer that says which edition it came from. A figure without an edition is not an answer. **Inputs** - `report`: a UN flagship report, e.g. `Human Development Report` - `figure`: what is wanted, e.g. `HDI for Ghana` ## Why this exists Flagship reports come out every year or two, and each edition recomputes past years. An assistant answering from memory tends to give the numbers of the edition it saw in training and call them current. In a recorded test, Claude without this skill named the 2025 Human Development Report and quoted the 2023/24 edition's Ghana figures. ## Sources (use these and nothing else) The publisher's own page for the report, then the data file linked from that page. Start here: | Report | Publisher | Where the edition is listed | |---|---|---| | Human Development Report | UNDP | `https://hdr.undp.org/data-center/documentation-and-downloads` | | Sustainable Development Goals Report | UN DESA | `https://unstats.un.org/sdgs/report/<year>/` | | World Population Prospects | UN DESA Population Division | `https://www.un.org/development/desa/pd/` | | State of Food Security and Nutrition in the World | FAO, IFAD, UNICEF, WFP, WHO | `https://www.fao.org/publications/fao-flagship-publications/the-state-of-food-security-and-nutrition-in-the-world/en` | | World Economic Situation and Prospects | UN DESA | `https://desapublications.un.org/publications/` | | World Health Statistics | WHO | `https://www.who.int/data/gho/publications/world-health-statistics` | | State of the World's Children | UNICEF | `https://www.unicef.org/reports/state-of-worlds-children` | | Employment and Social Trends | ILO | `https://www.ilo.org/research-and-publications` | | World Investment Report | UNCTAD | `https://unctad.org/topic/investment/world-investment-report` | | State of the Global Climate | WMO | `https://wmo.int/publication-series/state-of-global-climate` | For the HDI the data file is the composite indices time series CSV linked from the downloads page. Do not hard-code the file name. It carries the edition in its prefix, so read it off the page you just opened (checked on 2026-10-25: the file was named `HDR<YY>_Composite_indices_complete_time_series.csv`). The value is in column `hdi_<year>`, the rank in `hdi_rank_<year>`. ## Steps 1. Open the publisher's page. Find the newest edition and its publication date. Copy the line that shows both, word for word. 2. Follow that page to the edition's data file or report. Do not use a file from another edition, even for older years. Editions recompute the past. 3. Read the figure from the file. Note the column or table and the row. 4. Answer in this order: edition, publication date, figure, where in the file, source URL, date you read it. ## Answer layout ``` <report> <edition> (published <date>, "<line from the publisher's page>") <figure>: <value> <file>, <column or table>, <row> Source: <URL> · read <today's date> ``` ## Rules - **Cannot open the page or the file: say "I can't check the current edition" and give the link. Do not give a number.** - Never answer from memory, even with "about" or "I believe". A remembered figure is from an unknown edition. - The edition year is not the publication year. The 2025 State of the Global Climate came out in 2026. Copy both as the page writes them. - If you remember a different figure, say which edition you think it came from, then give the current one from the file. - Do not mix editions in one table. If a reader needs a time series, take every year from the same file. ## Copy-paste prompt (works in any assistant) > Before you give any figure from a UN report, tell me which edition it comes from and when that edition was published. Read the publisher's own page and the current data file first. If you can't open them now, say "I can't check the current edition" and give me the link instead of a number. Do not answer from memory. ## Running it as code `engines/material/` in the claude-for-sdg repository keeps a copy of each publisher page with its SHA-256 and accepts an edition or a date only when it is written in the quoted sentence of that stored page.