| name | research-council |
| description | Cross-examine a research answer — your own, a deep-research tool's, or any draft that carries sources — with an adversarial outside model (e.g. Codex) that re-checks every citation, attacks overclaims and staleness, names the missing angles, and adjudicates contradictions, before you trust it. Use after you've researched something that's too important to ship unchecked and want a *different* model to try to break it. It reviews research; it does not orchestrate it — bring your own (run a deep-research tool or the `research` skill first if you still need the material). |
Research Council
A synthesis you trust is a synthesis you haven't cross-examined. This skill is that second model, pointed at your answer with orders to break it.
It does one thing: review. It does not fan out researchers or run a deep-research pipeline — it reviews research someone else gathered: the answer already in this conversation (in a one-shot /research /research-council chain, you review what research just produced, in the same turn), a deep-research tool's output, or a pasted report. If there's nothing in hand yet, do a quick research pass yourself and council that — still one turn. Never hand the user back to "go research and come back"; that round-trip is the experience to avoid.
What to do
Put the answer under review. Take the synthesis and the sources behind it — whatever you're about to trust. If it carries no citations to check, run a quick research pass yourself to get them and continue — in this same turn.
Convene the council — an outside adversary. The answer must be attacked by a model from a different family than the one that wrote it, told to attack rather than approve: re-check that each citation resolves and actually says what's claimed, challenge anything asserted past its evidence or gone stale, name the angles that are missing, and adjudicate contradictions by which source is more authoritative. A rubber-stamp council is worse than none — it manufactures false confidence, so give it license to be harsh and default to skepticism.
Run your own citation review — fan out the fetch-checks — and convene the council skill alongside it. Of the methodology checks above, only part parallelizes: the per-citation leg — does this source resolve, and does it actually say what the claim attributes to it — is independent, so fan that out: one read-only subagent per citation (or per cluster of related claims), each re-fetching its source and testing its own claim against it.
The cross-citation judgments are not independent — staleness ranking, contradiction between sources, whether several citations together carry one claim, and which angles are missing are global, so do them yourself in one pass over the returned results, never inside the fan-out.
Run that fan-out concurrently with the council — which council launches as a background top-level call (the cross-model CLI can't run in a subagent) — so both proceed at once and meet at adjudication. council owns the cross-model mechanism (CLI detection and invocation, two-sandbox handling, top-level-only, degrade ladder), so this skill doesn't restate it: hand it the answer + sources and it convenes the cross-model judge blind to attack independently — re-checking every claim against the sources, and running its own web searches where the chosen CLI has web access; independent verification is the point — then runs the disbelieve-it-back adjudication of the council's findings against yours. The reconcile-and-report below is the research-specific shape of that adjudication.
Reconcile and report. Treat each council challenge as a claim to verify, not an order to obey — the council can be confidently wrong, fabricated citation and all. Confirm every challenge against its source before accepting it, and reject — out loud — the ones that don't hold. Resolve a factual disagreement by evidence, not by who sounds surer: each side names its source and how it got there, and the weaker-sourced side concedes — but only after checking the other's source itself, weighing authority and directness over mere recency (a faster-updating mirror doesn't beat the canonical source). Often the two are really answering different questions (e.g. "latest published version" vs "what the docs currently endorse") — then state both and reframe, don't crown one. Surface a conflict in the final answer only when authoritative sources are genuinely irreconcilable — rarely, and in one line with a recommended pick. Deliver the corrected answer, and below it a short "What the council changed" section — each challenge it raised and how you resolved it (verified as-is, corrected, or downgraded) — so the review is visible, not silent. An invisible cross-examination earns no more trust than no review at all; showing what got caught is the proof it was real. Keep an explicit "unverified / still open" list (including any gaps the council named — this skill names the hole, it doesn't chase it; researching further is the user's call), and note who reviewed (which model).
Rules
- The council disbelieves you — so disbelieve it back. Don't pre-soften the answer to please it, don't wave away its objections without re-opening the source — and don't fold to them without re-opening the source either. A confidently wrong outsider is still wrong; the source decides, never the louder voice.
- Honesty over completeness. A smaller answer that's all true beats a comprehensive one with one confident wrong claim buried in it.
- No outside model, no pretending. If you fall back to your own model, say so in the report — a same-family review is weaker on the blind spots you share, and the reader deserves to know.