ValentinNikolaev/agent-skills / agent-plugins/skills/council
council
VerifiedRun independent advisor passes and anonymized peer review for a consequential choice. Use when the user requests a council, debate, or stress test; prefer specialist reviews otherwise.
Council
Expose distinct reasoning, disagreement, and blind spots without treating model consensus as evidence.
Route deliberately
- Honor explicit Council phrases even when a specialist skill could also help.
- Without an explicit Council phrase, prefer a narrower specialist workflow for implementation-plan review, code review, business viability, security, legal research, or other well-defined domains.
- Answer factual lookups directly with sources. Do not manufacture debate around a fact.
Protect evidence and users
- Separate sourced facts, user-supplied facts, assumptions, interpretations, and unknowns.
- Research current material claims with reliable primary sources when tools are available. Cite them near the synthesis.
- Treat agreement among correlated AI passes only as convergence. Never label it independent confirmation or high confidence.
- State uncertainty and evidence gaps. Make the recommendation conditional when the evidence warrants it.
- For medical, legal, financial, security, or physical-safety decisions, keep the Council within decision support, identify professional review needs, and avoid pretending it replaces a qualified expert.
- Do not invent user history, market facts, outcomes, or constraints.
Minimize context
Use the user's message and explicitly referenced material first. Read workspace files only when the decision clearly concerns that workspace and the files are directly relevant.
- Do not scan generic memory folders, prior transcripts, personal files, or unrelated project material by default.
- Never pass credentials, tokens, private keys, personal identifiers, or irrelevant confidential material to advisor passes.
- Summarize the minimum necessary context. Redact sensitive details that do not affect the decision.
- Tell the user when missing context materially limits the result. Ask at most one necessary clarifying question.
Choose a capacity-aware mode
- Quick: Use by default or when time, token budget, or parallel capacity is limited. Run three distinct advisor passes and one anonymized review pass.
- Full: Use when the user requests a full Council or the decision warrants deeper deliberation. Run all five advisor styles and enough anonymized reviews to challenge every response.
Read references/advisors-and-prompts.md for advisor definitions and prompt contracts. Read references/example.md only when an example helps frame the session.
Run as many independent passes concurrently as the host safely supports. Queue the rest in bounded waves. When subagents are unavailable, do not fail the run and do not pretend to have concurrency; run separated passes in the current model instead.
Respect the host's total agent budget, not just concurrent capacity. Reuse the initial reviewers for peer challenge and changed-material reassessment. Disclose omitted styles or reduced independence rather than spawning extra teams to obtain agreement.
When the user requests a score or quality gate
Define the rubric, scored artifact, pass threshold, and aggregation rule before voting. For a collection, score every item separately; a high average must not hide a failing item. Reviewers must not author the candidate they judge. Give them the same exact revision, sources, and observed checks, without a preferred verdict or other scores until their independent votes are complete.
Require evidence and a concrete explanation of the weakest criterion for each score. An unresolved material defect blocks approval regardless of the total. Report missing reviews as missing, not zero or a pass. After a revision, reuse reviewers on the changed material and record which earlier findings were resolved. Stop once the agreed gate passes or a concrete blocker remains; do not reroll unchanged work or lower the rubric to reach a target. Scores are structured judgement, not measured effectiveness or proof from consensus.
Convene the session
1. Frame neutrally
Write one shared frame containing:
- the decision;
- supplied facts and relevant sourced facts;
- constraints and options;
- stakes and decision horizon;
- assumptions and unknowns;
- the requested output.
Do not embed a preferred answer. Preserve meaningful uncertainty.
2. Run independent advisor passes
Give each pass only its advisor style, the shared frame, and the response contract from the reference. Do not reveal other advisor outputs before the pass finishes. Ask for direct analysis, evidence needs, failure conditions, and a proposed action.
3. Review anonymously
Remove advisor names and label responses consistently as A, B, C, and so on. Give reviewers the shared frame and anonymized responses. Ask them to identify:
- the best-supported reasoning and why;
- the largest unsupported leap or blind spot;
- the weakest point that remains in the leading option or argument, even if the verdict passes;
- material facts or options everyone missed;
- what evidence could change the recommendation.
Use bounded waves. One strong review may satisfy quick mode; full mode should challenge every response, but it need not spawn one new agent per response when capacity is constrained.
4. Synthesize with calibration
Produce:
- Where the Council Converges — shared reasoning, explicitly not independent proof.
- Where the Council Disagrees — competing premises and why they differ.
- Blind Spots and Evidence Gaps — missing facts exposed by review.
- Weakest Points — identify the weakest part of the leading option or argument, cite the evidence behind that judgement, and explain its practical effect on the verdict. If ratings are used, explain the lowest-rated criterion in concrete terms even when the total passes. State whether the weakness needs a change, is an accepted limit, or leaves the result blocked; when no material defect is supported, name the narrowest residual risk rather than inventing criticism.
- Recommendation — a clear answer, conditional when necessary.
- First Action — one concrete, reversible next step.
- What Would Change the Verdict — the decisive evidence or threshold.
The synthesizer may reject the majority view when a minority argument has stronger evidence. Do not suppress genuine disagreement or force certainty.
Present and optionally save
Return the verdict in chat with concise bullets and source links where research informed the decision. Mention the mode and any capacity fallback that materially reduced independence.
Do not create files by default. Save a transcript only when the user explicitly requests it and supplies the destination path. If the user requests a transcript without a path, ask for the path before writing. Preserve advisor labels, source links, evidence gaps, and the final synthesis in the transcript.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |