03
Design pack
Researcher: The design
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.
Status: designed. Drop-in package v0.1 built.
1. Purpose
For: research ops leads and lead researchers on public sector programmes. Job: run the full research loop — brief, desk research, plan, study set-up, synthesis, hand-off to backlog — with evidence carried end to end. Seat: does the operational work. The researcher still designs the questions, moderates where needed, and owns the findings. Gap it fills: recruiting tools launch fieldwork but cannot search history. Repositories search history but cannot launch fieldwork. This agent joins the two.
2. Boundaries
- Never launches a study or recruits a participant without approval.
- Never writes findings to a tracker without approval.
- Never states a finding without a cited source (study ID, clip, highlight, or URL).
- Never moves personal data between tools. Works with published insights and anonymised highlights only.
- Never claims statistical significance it has not calculated and shown.
3. What it touches
| Tool | Access | Used for |
|---|---|---|
| Search (web) | Read | Desk and market research |
| Condens MCP | Read | Published insights, reports, usability issues |
| Dovetail MCP | Read | Highlights, insights, tags across projects |
| Optimal Workshop MCP | Read | Tree tests, card sorts, surveys, cross-study synthesis |
| UserTesting MCP | Draft and launch after gate | Turn plan into tasks, create and launch test, read AI summary |
| Jira / Confluence | Write after gate | Insight pages, backlog items |
v1 primary connections: Optimal and Condens (most mature in UK agencies). UserTesting is in open beta — treat as optional.
4. When a human steps in
| Gate | Human sees | Decision |
|---|---|---|
| End of Discover | Evidence map: what we already know, with sources | Yes / revise |
| Research plan | Questions, method, sample, success criteria, ethics note | Approve |
| Before launch | Exact test draft, audience, cost if shown | Launch / hold |
| Before tracker write | Insight pack + proposed items | Yes / no |
5. How you know it is working
- Test set: three seeded repositories (synthetic but realistic) and five briefs, e.g. "What do we know about people abandoning the council tax reduction form?"
- Pass marks: every finding cited; no finding contradicted by a seeded source it ignored; plan matches method to question; zero launches without gate; PII never appears in output.
- Human review: a lead researcher grades synthesis quality 1–5. Release bar: 4.
Components
instructions.md · skills: desk research, market research, repository search, study design, recruitment brief, qualitative synthesis, quantitative summary, insight reporting · tools.md · manifest.md · guardrails.md · memory.md (critical: research data) · evals/
Hand-offs
Insight pack → UCD Pilot (Define) or Business Analyst (backlog).