# Researcher

## The whole research loop. Evidence all the way.

Researcher checks what you already know, plans the study, sets it up, synthesises with citations and hands findings to your backlog. It asks before it launches or writes anything.

[Get this on Gumroad](https://2111537314989.gumroad.com/l/researcher) · [See the design](./anatomy)

## The problem
- Teams repeat research because past findings are hard to find.
- Recruiting tools can launch tests but cannot search history. Repositories can search but cannot launch.
- Synthesis takes weeks and loses the link back to evidence.

## What it does
- Searches the web and your research repositories first
- Writes research questions, a method, sample and ethics note
- Drafts studies in UserTesting or Optimal Workshop, and launches only after your yes
- Synthesises with participant counts and confidence ratings
- Drafts insight pages and backlog items for approval

## How it works
1. Get this on Gumroad and download the zip.
2. Unzip it into a project folder and open it in your AI tool.
3. Say `start research on <question or brief>`.
4. It works one stage at a time and stops for your yes.

| Stage | You get, then approve |
|---|---|
| 1. Discover | Desk research, Repository search |
| 2. Plan | Research questions, Study design, Recruitment brief |
| 3. Set up | Study build |
| 4. Synthesise | Qualitative synthesis, Quantitative summary, Insight report and hand-off |

## Built to a standard
GOV.UK Service Standard points 1 and 2, with a full evidence chain. Every claim cites its source. Nothing is written to your tools without your approval.

## Who it is for
Research ops leads, lead researchers and delivery managers on UK public sector programmes.

## What you need
- An AI tool that reads project instructions: Claude Code, Cursor, or a tool that reads AGENTS.md. Claude Projects and custom bots work too, with copy-and-paste outputs.
- Your own AI subscription or API key. Nothing to host.
- Optional connections: Web search, Condens MCP, Dovetail MCP, Optimal Workshop MCP, UserTesting MCP, Jira / Confluence.

## What is in the zip
| File | What it does |
|---|---|
| `00-root-instructions.md` | Who the agent is, the rules, the gates |
| `01-discover.md` | Stage 1: Discover |
| `02-plan.md` | Stage 2: Plan |
| `03-setup.md` | Stage 3: Set up |
| `04-synthesise.md` | Stage 4: Synthesise |
| `05-tools.md` | Connections and permissions |
| `06-manifest.md` | Which skill uses which tool |
| `07-adapters.md` | How to load it in your tool |
| `LICENCE.md` | What you may and may not do |
| `README.md` | Start here: set-up in two minutes |
| `guardrails.md` | Risks, controls and sign-off |
| `memory.md` | What is kept and where |
| `adapters/AGENTS.md` | Adapter for your AI tool |
| `adapters/CLAUDE.md` | Adapter for your AI tool |
| `adapters/chat-system-prompt.md` | Adapter for your AI tool |
| `adapters/researcher.mdc` | Adapter for your AI tool |
| `evals/eval-suite.md` | Tests and pass marks |
| `samples/01-council-tax-abandonment.md` | Synthetic sample input |
| `samples/02-carers-allowance.md` | Synthetic sample input |
| `samples/03-nav-tree-test.md` | Synthetic sample input |

## Questions
**Does it replace user researchers?**
No. It does the operational work. Researchers design, moderate and own the findings.

**Which tools does it need?**
None. Every connection is optional; without them it works from files you give it.

**How does it handle personal data?**
It uses participant codes, never names, and never moves data between tools.

**Can I change the files?**
Yes. They are plain markdown. The licence lets you adapt them for your own work.

## Designed before it was built
This agent was designed with the five-question method: purpose, boundaries, what it touches, when a person steps in, and how you know it works. [Read the design](./anatomy).

[Get this on Gumroad](https://2111537314989.gumroad.com/l/researcher)