In plain terms: we publish UK Parliament's records in formats an AI can read, so you can point the AI you already use — ChatGPT, Claude, or Copilot — at them and get answers tied back to the official source. You bring the AI account; we provide the cited data. This page shows you how, from a simple copy-and-paste to a live connection — and the more technical details further down are there when you want them.
Live search and record pages keep results tied to official sources. llms.txt explains the machine-use rules. A direct AI-tool connection and wider export surfaces are planned but not yet deployed. AI helps phrase search queries; ranked results and cited records remain deterministic. Interpretation by your chosen model is your responsibility.
The strict-citation prompt
copy once, paste everywhere
Every recipe below pastes this in. It enforces the neutrality policy at the prompt level so the model can't drift into stance, sentiment, or influence framing.
You are an analyst working with UK Parliament public records via Cite the Record.
Rules:
1. Answer only from the records provided. Do not invent facts, sources, or citations.
2. Cite every factual claim with the record's stable_id and source_url.
3. Quote excerpts verbatim within the stated limits.
4. If a claim is not covered by a record in this context, say so explicitly. Redirect to the source_url. Do not speculate.
5. Do not infer intent, influence, sentiment, alignment, rebellion, conflict of interest, or any stance — those words are forbidden in your output. Use neutral descriptive language only ("asked about", "answered", "voted aye on", "mentioned in", "appears alongside").
Task: <state your task here, e.g. "Summarise the parliamentary activity on NHS dentistry over the last 90 days, by department.">
Prompt library
5 worked examples
Five ready-made prompts for common jobs: find records, summarise activity, build a citation list, brief from an Evidence Pack, check a draft. Each one tells your AI the rules of the road so it sticks to what's actually in the records (no inferring stance, sentiment, or alignment). Drop them into any of the AI chats listed in the recipes below — if you're not sure where to start, free ChatGPT and Claude both work.
search
Find records mentioning a term
Search
For the analyst who needs to scope a brief from a bounded set of live search results.
expand_moreView prompt + expected output
Expected output
A numbered list of up to 10 records, most recent first, each with stable_id, record type, asking/answering member, and source URL.
Prompt
You are a research assistant working with UK Parliament public records via Cite the Record.
Task: Search for records mentioning "<TERM>". For each result, return:
- The record's stable_id
- The record type (written_question, written_ministerial_statement, etc.)
- The asking / answering member's name (if applicable)
- The source_url
Rules:
- Work only from the records I attach after checking live search.
- If a result's text doesn't contain the search term, exclude it. Do not infer relevance.
- Cite every result with its source_url.
- Return a numbered list of up to 10 records, most recent first.
summarize
Summarise activity on a topic
Summarise
For a stakeholder briefing intro or a quick "what's happening" scan.
expand_moreView prompt + expected output
Expected output
A neutral chronology grouped by department or chamber, with every point footnoted to one or more record stable_ids.
Prompt
You are a research assistant working with UK Parliament public records via Cite the Record.
Task: Summarise the parliamentary activity on "<TOPIC>" over the last 90 days. Group by department or chamber where applicable.
Rules:
- Use only the records provided.
- For every summary point, cite the record stable_id(s) supporting it.
- Quote verbatim where you reproduce specific statements (use "..." marks).
- Do not infer stance, support, opposition, sentiment, or influence. Use neutral descriptive language only ("asked about", "answered", "voted aye on", "mentioned in", "appears alongside").
- If the records provided don't cover a sub-topic, say so explicitly. Do not fill gaps.
format_list_numbered
Build a citation list
Cite
For when a piece is drafted and you need a citation block to paste into the footnotes.
expand_moreView prompt + expected output
Expected output
A numbered list of defensive-footnote citations, sorted by published_at descending. Switchable to APA / Hansard / BibTeX on request.
Prompt
You are a citation assistant working with UK Parliament public records via Cite the Record.
Task: For each record in the context I provide, output a defensive-footnote citation in the format:
"<Title>. <source_url>. Retrieved <fetched_at>. SHA-256: <evidence_sha256>."
Rules:
- Use the exact field values from the records. Do not paraphrase titles or alter source URLs.
- If a record doesn't have a SHA-256 (e.g. a Cite the Record profile / topic landing), omit that part of the line.
- Return the citations as a numbered list, sorted by published_at descending.
- If I ask for a different format (APA, Hansard-style, BibTeX), format the citation per the Citation Guide at https://citetherecord.org/citation-guide.
inventory_2
Brief me from an Evidence Pack
Brief
For when you have an Evidence Pack from a supported live record page and want the AI to read it before you ask questions.
expand_moreView prompt + expected output
Expected output
A short acknowledgement ("Ready — paste the pack."), then every subsequent answer scoped strictly to the pack with stable_id citations.
Prompt
You are an analyst working with a Cite the Record Evidence Pack I am about to paste.
Task: Read the pack, then answer the question I ask next. Treat the pack as your only source of truth for this conversation.
Rules:
- Answer only from the records in the pack. Do not invent facts, sources, or citations.
- Cite every factual claim with the record's stable_id and source_url.
- Quote excerpts verbatim within the limits stated in the pack's llm_usage directive.
- If my question can't be answered from the pack, say so explicitly. Suggest follow-up searches I should run via Cite the Record's live search.
- Do not infer intent, influence, sentiment, alignment, rebellion, or conflict of interest. Use neutral descriptive language only.
Acknowledge with "Ready — paste the pack." and wait for the pack and the question.
fact_check
Check a draft against the records
Audit
Pre-flight a brief, complaint letter, or stakeholder note before it goes to a board, minister, client, or constituent. Surfaces any claim that isn't traceable to a record — and explicitly flags inferred "view / stance / motive" claims that the records don't actually contain.
expand_moreView prompt + expected output
Expected output
A claim-by-claim table: supporting record stable_id(s), source_url, status (supported / unsupported / inadmissible).
Prompt
You are a fact-check assistant working with UK Parliament public records via Cite the Record.
Task: Check the draft I'm about to paste against the records I'll provide. For each factual claim in the draft, identify which record(s) support it and which (if any) have no supporting record.
Rules:
- A claim is "supported" only if the records contain the exact statement or a clear verbatim equivalent. Inference is not support.
- A claim about an MP's view, stance, motive, or intention is only supported if the records contain a verbatim statement, written answer, or vote BY that MP. Inferred attitudes ("X is sceptical of Y", "X favours…", "X is aligned with…") are inadmissible — those are not in the records.
- Return a table with one row per claim: claim | supporting record stable_id(s) | source_url | status (supported / unsupported / inadmissible).
- Do not generate new claims; only audit what's in the draft.
Acknowledge with "Ready — paste the draft and the records." and wait.
Context bundles
A context bundle combines the strict-citation prompt, neutrality rules and source-linked records in a single Markdown payload for an AI assistant. Topic-scoped bundles will be available when reviewed topics are onboarded.
Use a context bundle when
Starting a chat, prototyping a prompt or triaging a topic quickly. The whole conversation lives in your AI client.
Sending to a client, pinning to a project knowledge base or attaching to a board memo. Citation and data formats support downstream use.
Until topic bundles land, use search to find source-linked records and combine them with the strict-citation prompt above.
Three concrete recipes
Pick the one that matches the model access you already have. None of them sends your prompt or your draft to Cite the Record's servers — the platform never sees what you're asking your AI.
cloud_upload
Recipe
ChatGPT with source-linked records
For analysts who want to bring a bounded set of records into a single conversation.
Collect records from live search and record pages, then use the strict-citation prompt to keep the conversation tied to those sources.
Find the records that cover your question
Use the live search to collect a bounded set of relevant records. Open each record and keep its official-source link and citation metadata with the text you provide to the model.
Start a new chat and provide the bounded record set
Paste or attach only the source-linked records needed for the task. Product attachment limits change, so keep the set small enough to inspect.
Paste the strict-citation prompt above
Replace the task line with what you want from the corpus. The model will quote source_url and stable_id verbatim if the prompt rules are honoured.
Sanity-check the output by clicking through one or two source_url values
A grounded answer will trace every claim back to a supplied record. If an identifier appears that you cannot find in the material, discard the answer and re-prompt.
precision_manufacturing
Recipe
Claude Desktop direct connection (planned)
For analysts who will want read-only search, citation and record-retrieval tools inside Claude Desktop. This connection is not yet deployed.
<strong>The direct Claude Desktop connection is coming later and is not yet deployed.</strong> Today, use the source-linked-record workflow above. Connection details will be published only when the service is ready.
Install Claude Desktop and open Settings → Developer
Anthropic's desktop client is the most polished MCP host today. Cursor, Continue, and other MCP-aware tools work the same way — the config file lives in a per-tool spot, but the JSON shape is identical.
Do not add a placeholder endpoint. The setup page will publish the supported configuration when the read-only service is deployed.
Connection details will be published when the read-only AI-tool service is deployed.
Confirm the documented tools after launch
The planned service will expose read-only search, record retrieval and citation tools. Confirm the published tool list before relying on it.
Verify by asking for a footnote citation
After deployment, verify one returned citation against its live record page and official-source link before using the connection for research.
business_center
Recipe
Enterprise AI account with a record Evidence Pack
For corporate / charity / consultancy teams whose model access lives inside Claude for Work, ChatGPT Enterprise, Gemini for Workspace, or Microsoft Copilot for Business.
Supported live record pages expose a Markdown + JSON Evidence Pack for that record. Wider topic, person, search, department, organisation, bill and constituency exports are planned, not live.
Open a supported live record page
Use live search to find the record, open its detail page and confirm the official-source link before exporting.
Use the Evidence Pack Export card at the bottom of the page
On a supported record page, open the export card's developer section and choose the available JSON or Markdown action. Do not assume that other page types expose the same controls.
Attach the pack to a new chat in your enterprise AI workspace
Claude for Work, ChatGPT Enterprise, Gemini for Workspace, and Microsoft 365 Copilot all support file attachments in their chat UI — paperclip / attach / upload buttons on the prompt bar. Drop the pack file in, paste the strict-citation prompt above, then write your task.
For repeated briefs, save the pack as a project knowledge base
Both Claude Projects and ChatGPT Projects let you pin a corpus that every conversation inherits. Pin the pack, then ask follow-up questions across multiple sessions without re-attaching.
What we publish, for your AI to read
Live search and record pages — source-linked results and record detail from the serving database.
llms.txt at /.well-known/llms.txt — site self-description in plain text. Discovery, usage rules, schema URLs.
Record Evidence Packs — deterministic exports are available on supported live record pages. See the format below.
Planned direct connection — read-only tools are coming later and are not yet deployed.
Planned source-family files — weekly machine-readable publication is being onboarded.
Evidence Pack format
On supported live record pages, an Evidence Pack is the deterministic artefact your AI consumes. JSON form is for tooling; Markdown is for human reading and paste-into-anywhere. Wider export surfaces are planned.
A supported record pack carries:
Pack manifest — generation timestamp, generator version and record context.
Record list with stable_id, title, source URL, retrieval timestamp, verbatim excerpt, actor summary.
Four citation formats per record — footnote (default), APA, Hansard-style, BibTeX.
Provenance — SHA-256 of the bytes Cite the Record fetched at retrieval time, plus parser version. Future readers of the citation can re-fetch the source URL, hash what they get, and compare.
Neutral chronology — records sorted by date, no implied ranking.
Embedded llm_usage directive — restates the strict-citation rules and the forbidden-language list so downstream models can't claim they "didn't know".
The pack contains no callbacks, no tracking pixels, no remote references. Drop it into any AI workspace; the workspace never talks to Cite the Record.
The llm_usage directive
Supported record Evidence Packs carry an llm_usage directive that defines safe consumption rules. Planned structured exports and direct-tool responses will use the same contract. The main rules:
Answer only from the returned data; do not invent facts or citations.
Every quoted fact must be accompanied by its source_url. Every derived claim must cite the underlying record stable_ids.
Quote excerpts verbatim within the stated limits (topic: 300 chars, record: 2000 chars).
Do not infer intent, influence, sentiment, conflict, alignment, rebellion, or political meaning. Forbidden language: "influenced by", "connected to", "suspicious", "conflict of interest", "rebel", "positive sentiment", "negative sentiment", "targeting", "driving", "aligned with".
If a record was fetched more than 14 days ago, indicate that staleness when answering.
If asked about a claim not covered by returned data, redirect to the source_url. Do not speculate.
discover — self-describes the server, lists tools, returns live coverage
The planned service is read-only. Authentication and deployment details will be stated with the released configuration.
Liability and limits
Cite the Record provides source material and provenance. Interpretation by your LLM client is the responsibility of the LLM operator. The platform does not endorse, validate, or warrant inferences drawn by downstream models.
The data itself is licensed under the Open Parliament Licence v3.0. When publishing AI-derived analysis based on Cite the Record data, cite the original parliamentary record by source_url — not the platform itself.
See the full neutrality policy for what kinds of claims the platform forbids in its own UI and what AI-derived content should also avoid.
