| name | research-review |
| description | Review a research effort for grounding, internal consistency, and synthesis fidelity before its conclusions are relied on — and to catch the contradictions that iteration introduces across question files and the synthesis. Reports findings; does not auto-fix unless asked. |
Research Review Skill
Review research the way a skeptical reader would: confirm the claims are grounded, the question files don't contradict each other, and the synthesis faithfully represents what the questions actually found. This is the reactive net for the drift that creeps in as research is iterated (the proactive side is the propagate-on-edit rule in research-resume).
When to use
- Before a research synthesis is relied on (to feed a plan, an ADR, or a decision).
- After the research has been iterated — a review applied, a late finding added, an owner direction folded in — the moment most likely to have left one file contradicting another.
- When asked to "review" or "sanity-check" research.
What to check
Verify each load-bearing claim against the actual files and source before reporting it.
- Internal consistency (the drift net — the highest-value pass).
- No contradictions across question files. A decided value, rule, or vocabulary term stated in one question must read the same in every other question that references it. (The classic failure: a rule is sharpened in one file but its echo in another file's takeaways/table still states the old form.)
- The synthesis faithfully represents the findings. Every claim in
synthesis.md traces to a question file; the synthesis does not invent, overstate, or quietly resolve a contradiction the questions left open. Where questions disagreed, the synthesis says so (or the disagreement was reconciled in the questions first).
- No stale references. Grep across all question files and the synthesis for any value changed during iteration (a renamed term, a superseded recommendation, an old number) and reconcile every occurrence.
- Grounding. Claims are cited to source —
file:line for codebase facts, precise citations (RFC numbers, spec sections, doc URLs) for prior art. Load-bearing claims are actually verifiable, not asserted.
- Completeness. Each investigation question is genuinely answered (not deflected); open questions are surfaced rather than buried; recommendations are supported by the findings, with a clear recommendation per decision rather than a survey.
How
- Read every question file and the synthesis. Build a small map of the shared terms/decided values and check each one reads consistently everywhere it appears.
- For each value changed in iteration, grep all files for the old form. For each synthesis claim, confirm it traces to a question.
- Spot-check load-bearing source citations against the actual source.
- Ground every finding in a specific file + location and state the minimal remedy.
Output
Report findings grouped by severity, each with location and a minimal fix — do not rewrite the research unless explicitly asked to apply fixes:
**[Blocking]** <contradiction / unsupported claim> — `q5-…md:NN` vs `q2-…md:NN`. <why it matters> Fix: <minimal reconcile>.
**[Should-fix]** …
**[Nit]** …
Close with a one-line verdict (sound / sound-with-fixes / needs-reconciliation). If asked to apply the fixes, make the minimal change and re-sweep all files (per research-resume's propagate-on-edit rule) so the fix propagates everywhere the value appears.
