expert-panel

A good brainstorm is several experts, each master of a domain, who bring their knowledge and argue toward a decision. Use this when a choice has no ground truth to check against yet — a design fork, an architecture call, a tradeoff. For anything that can be checked against an oracle, use servo-gate instead; this skill is for generating and stress-testing options, not verifying them.

The failure this prevents: one model picking the first plausible option with false confidence, and a panel of same-context models agreeing with each other.

Keep it visible: open a TodoWrite list with one todo per expert (plus the dissenter and the synthesis), and mark each in_progress/completed as it reports. The human should see the panel convene and resolve, not wait on a silent block.

Procedure

1. Frame the fork. State the decision and the constraints from the context brief (gather-context). If there is genuinely an oracle that settles it, stop — this is a gate, not a panel.

2. Convene the MINIMAL relevant set — usually 2-3, not every plausibly-related domain. From the manifest's experts, pick only the ones the decision genuinely turns on (a data fork wakes the data expert; it does not need the bot expert). Convening six experts on a three-expert problem is the waste this step must avoid.

   How to spawn an expert: an expert is not a named agent type — do not try to spawn an agent called data or hs-coaching (it does not exist and will fail). Spawn a plain general subagent and give it the expert's identity as context: the expert's sources from the manifest (read any skill: pointer as a file, read the listed code/notes), plus the context brief from gather-context, plus its assigned stance. That content IS the expert. Hand it the brief so it does not re-read the whole project — it only goes deeper into its own slice. Each expert is a real, separate perspective — not you role-playing all of them. These reader-experts can run on a cheaper, faster model; reserve the strong model for the synthesis in step 5.

3. Seat the mandated dissenter. Spawn one more perspective whose assigned job is to disagree with the emerging consensus — by duty, not by opinion (the "tenth man"). If everyone is converging, it must surface the strongest case against. Keep its dissent on the record even if overruled.
4. Place the human. The human is the CEO/client seat: they carry the intent and make the call. If they also hold a domain, their input is an expert contribution — and gets challenged like any other, not executed blindly.
5. Synthesize, do not average. Present the options scored, with the dissent kept visible. Recommend, but do not collapse to a forced consensus. Propose the open choices to the human in one batch, pre-filled.

Scale to the stakes (cost)

The panel is the most expensive move in the flow — each expert is a subagent. Scale it:

• Default (most decisions): 2-3 experts, on a cheap model, fed the brief. Often a single dissenter is enough alongside them.
• High-stakes only (irreversible, expensive to reverse, security): widen the panel and run experts on the strong model.

If the human or the flow signalled a quick pass, lean to the smaller end (even reason from 2 lenses inline rather than spawning, accepting less independence for a low-stakes fork). A full panel on a routine fork is the ceremony this skill must not become.

Rules

• Experts must not collude: each sees its own slice and stance, not the others' reasoning, until synthesis.
• The dissenter is structural. A panel with no dissent is the groupthink this skill exists to break.