Wiki · Agent templates
Build with the catalog
Agent templates
Pre-built setup instructions for wiring Claude Project, Custom GPT, Gemini Gem, NotebookLM, or Microsoft Copilot to query Policy Window's primary-source-cited AI-governance catalog. Each template ships a system-prompt template + a list of knowledge files to attach + sample queries + charter-alignment notes. Released under CC0 1.0: fork, remix, redistribute, no attribution required.
Who these are for: AI-governance teams, journalists on deadline, procurement evaluators, regulator staffers, civil-society advocates, and developers prototyping retrieval flows — anyone who wants a citation-defensible conversational front-end onto the catalog without writing the system prompt from scratch.
Pick your platform
Claude Project
Claude.ai Projects with Custom Instructions + uploaded knowledge files. Best agent for verbatim-quote fidelity + structured citation envelopes; the only template here with optional MCP wiring.
Setup guide →
Custom GPT (OpenAI)
Deprecation announced 2026-02ChatGPT Custom GPT with system instructions + Knowledge attachments. Limited to attached files; OpenAI announced a deprecation track in Feb 2026, so we recommend planning a migration to Claude Project or Gemini Gem.
Setup guide →
Gemini Gem
Google Gemini Gem with system instructions. Best agent for cross-language coverage when querying non-English primary sources; Google search grounding helps surface PW pages.
Setup guide →
NotebookLM
Notebook bundle ships iter-320Google NotebookLM with PW source pack attached. Best agent for source-grounded summarisation + citation-by-passage; the per-source citation pop-out enforces verbatim-quote discipline by default.
Setup guide →
Microsoft Copilot / Bing / Edge
Microsoft Copilot in the Edge sidebar + Bing-grounded conversational mode. Best agent for end-users already inside the Microsoft ecosystem; works via web grounding rather than knowledge-file uploads.
Setup guide →
What about MCP (Model Context Protocol)?
Policy Window publishes an MCP descriptor at https://policywindow.org/mcp.json documenting the 8 read-only tools (search_catalog, get_instrument, get_topic_coverage, get_coverage_cell, compare_jurisdictions, list_changes, cite_article, glossary_lookup). MCP gives you tool- calls instead of bulk file uploads — agents fetch only what each query needs, citations land pre-formatted, and the response envelope carries the attribution string verbatim.
As of late 2025, MCP support is uneven across the platforms above:
- Claude Project / Claude Desktop:MCP supported today via the desktop client's MCP server configuration. See the Claude Project template for the wiring.
- Custom GPT: no MCP support as of late 2025; stick to Knowledge files + Actions.
- Gemini Gem: no MCP support as of late 2025; stick to system instructions + grounded web search.
- NotebookLM: no MCP support; source-based ingestion only.
- Microsoft Copilot: no consumer-facing MCP support as of late 2025; preview MCP support has been demonstrated for Copilot Studio (enterprise) but is out of scope for these templates.
The npm package @policywindow/mcp-server (v0.1.0-alpha) packages the 8 tools as a stdio MCP server for any MCP-compatible host. Install:
npm install -g @policywindow/mcp-server
v0.1.0-alpha; SDK install required separately, see /wiki/agents for live status + the canonical install path.
What these templates won't do
- Replace the catalog's editorial limits. The agent is bound by the methodology §11 disclosed limits (one editor of six, English-only primary sources, sectoral blindspots, etc.). The template wires the agent to surface those limits when asked.
- Provide legal advice. Per charter §7.4, every template instructs the agent to defer interpretive legal questions to qualified counsel.
- Take actions on the user's behalf. The agent is read-only retrieve-and-cite (per charter §7.1.b); no submit-comment, no post-to-regulator, no subscribe-to-alerts.
- Generate persuasive content. Per charter §7.6, every template prohibits the agent from producing personalised, microtargeted, or covertly persuasive output.