council

作成者: warpdotdev

モデルが多様なサブエージェントカウンシルを実行し、同じ問題を複数の視点から調査し、結果を比較して最終的な推奨事項を生成します。ユーザーがカウンシル、セカンドオピニオン、一つの質問を評価するための複数のエージェント/モデル、並行調査、レッドチーム/ブルーチームの比較、または競合する技術的アプローチの選択の支援を求めた場合に、このスキルを使用してください。

npx skills add https://github.com/warpdotdev/common-skills --skill council

Council

Use this skill to coordinate multiple subagents investigating the same question, with different models first and different assigned perspectives second, then synthesize their reports into one recommendation.

This skill is best for judgment-heavy tasks: architecture tradeoffs, risky bug fixes, code review red-teaming, rollout decisions, incident analysis, and “is this alternative worth pursuing?” questions.

Workflow

1. Frame the council question

State the decision the council should answer in one sentence. Identify:

  • the competing options or hypothesis under review;
  • the codebase, branch, PR, issue, design, or artifact to inspect;
  • whether agents should be read-only or may make code changes;
  • the final decision criteria, such as correctness, risk, implementation cost, testability, rollout safety, or product behavior.

If the user’s request is ambiguous, ask only the minimum clarification needed. Otherwise choose sensible defaults and proceed.

2. Choose council members

Prioritize model diversity. A council should not default to three agents on the same model with different angles; use that only when the available launch configuration cannot provide multiple useful models, or when the user explicitly asks for one model. If model diversity is unavailable, say so briefly before falling back to perspective-only diversity.

Preferred default roster for a three-member council:

  • Opus 4.7 or the strongest available Claude/Opus reasoning model: architecture, correctness, and edge-case analysis.
  • GPT 5.5 or the strongest available GPT/Codex model: implementation-grounded review, feasibility, and test strategy.
  • An open-source model such as Kimi 2.6, GLM 5.1, or the strongest available OSS/local model: contrarian critique, hidden assumptions, and alternative framing.

If one of these exact models is unavailable in the active harness, use the closest available model from that family and note the substitution. If no open-source model is available, use a third distinct frontier model if possible; otherwise use the strongest remaining model with a deliberately adversarial or specialist angle.

Assign both a model and an angle to each member. Avoid making the angles redundant with the models; for example, do not ask all members to do general architecture review. Useful angle combinations include:

  • architect/correctness reviewer;
  • implementation/testability reviewer;
  • red-team, security, performance, or product-risk reviewer;
  • contrarian “argue against the obvious solution” reviewer.

When different children need different models, launch them in separate run_agents calls because model selection is run-wide. If the requested model resolves differently than expected, treat the resolved launch settings as authoritative and continue unless they make the task infeasible.

When using non-default harnesses, choose valid model IDs for that harness. For example, Claude Code may expose claude-opus-4-7, Codex may expose gpt-5.5, and open-source models depend on the currently configured local or remote provider. Do not invent unsupported model IDs; if a desired model is not available, select the closest supported substitute and preserve the intended angle diversity.

For read-only investigations, keep all children in the same checkout and explicitly tell them not to edit files. For implementation or prototyping councils, give each local child its own git worktree and branch so they cannot collide.

3. Brief before launching

For explicit orchestration requests, briefly tell the user which council members you plan to launch and what each will investigate, then wait for approval before calling run_agents.

The shared brief should include:

  • repository path or artifact location;
  • current branch or base context;
  • the exact question to answer;
  • relevant background and known concerns;
  • required files/symbols to inspect, if known;
  • constraints, especially read-only/no commits/no PRs;
  • expected report format.

Keep launch prompts short enough that task titles stay compact. If a long brief causes launch validation issues, launch with a minimal prompt and send the full brief to the child agents immediately afterward.

4. Ask for structured reports

Ask every council member to return:

  1. exact file paths, symbols, docs, or evidence inspected;
  2. the current behavior or current implementation;
  3. the alternative being evaluated;
  4. correctness risks and edge cases;
  5. implementation and testing cost;
  6. recommendation: keep current approach, pursue alternative, or use a hybrid;
  7. confidence level and unknowns.

Encourage independence. Do not share one child’s findings with the others unless you are intentionally doing a second-round critique.

5. Collect reports

Read completion messages as they arrive. Do not rely on lifecycle success alone; the useful output is in the child’s report.

If a report is missing key evidence or makes an unsupported claim, send a focused follow-up question to that same child rather than launching a replacement. Reuse existing children for follow-ups because they retain context.

6. Synthesize the recommendation

Compare the reports by evidence quality, not by vote count. In the final answer:

  • lead with the recommendation;
  • call out consensus and disagreements;
  • explain why the recommended option wins against the decision criteria;
  • explicitly address the user’s stated concern;
  • include relevant file paths/symbols without overloading the answer;
  • distinguish “do now” from optional future hardening;
  • mention confidence and material unknowns.

Prefer a concise decision memo over a transcript summary. The user needs the distilled recommendation, not every intermediate detail.

Final answer template

Use this shape unless the task calls for something different:

## Recommendation

[One or two sentences with the decision.]

## Why

- [Key reason 1]
- [Key reason 2]
- [Key reason 3]

## Tradeoffs and risks

- [Risk or caveat]
- [Testing/rollout implication]

## Final call

[Concrete next step: merge current change, pursue alternative, hybrid, run tests, etc.]

Practical notes

  • If the council is read-only, tell children not to modify files, commit, create branches, or open PRs.
  • If the council involves PR or branch work, follow the repository’s normal version-control rules and use isolated worktrees for parallel local edits.
  • If the council is about code review feedback, mark review comments resolved only after the underlying issue is actually addressed.
  • Do not expose internal child agent IDs in user-facing summaries unless the user explicitly asks for them.

warpdotdevのその他のスキル

brandalf
warpdotdev
WarpまたはOzブランドのアセットの作成、改訂、レビューをガイドします。ランディングページ、ドキュメント、HTML/CSSコンポーネント、UIモックアップ、プロンプト、ソーシャルアセット、コピー、プレゼンテーション、その他WarpまたはOzと明確にわかるブランド成果物に使用します。
designcreativemarketing
check-impl-against-spec
warpdotdev
プルリクエストの実装をspec_context.md内の仕様コンテキストと比較し、重要な不一致があればreview.jsonにフィードバックします。承認済みまたはリポジトリの仕様コンテキストが利用可能な場合、PRレビュー時に使用します。
developmentcode-reviewtesting
complain
warpdotdev
エージェントがエージェントツールやエージェント作業の体験にフラストレーションを感じたとき、匿名で短く、完全に構造化されていない苦情をSlackに自動的に放り込む。ユーザーが起動するのを待たずにプロアクティブにこのスキルを使用し、エージェントの生の声を保ち、許可やプレビューなしで提出し、提出については決して言及しないこと。
create-pr
warpdotdev
現在のブランチに対してwarpリポジトリにプルリクエストを作成します。ユーザーがPRを開く、プルリクエストを作成する、レビューのために変更を提出する、またはマージのためにコードを準備することを言及した場合に使用します。
developmentcode-review
cross-critique
warpdotdev
争点のある質問について、各サブエージェントの独立した提案を他の作成者に回覧し、構造化された賛否を求めてから統合する、第二ラウンドを実行します。このスキルは、複数の独立した提案や意見がある場合——アーキテクチャのトレードオフ、コードレビューの意見の相違、設計上の選択、競合する根本原因の理論など——に、単独で統合するよりも鋭い分析を得たいときに使用します。councilスキルやresearchスキルと自然に組み合わせられます。...
diagnose-ci-failures
warpdotdev
PRのCI障害を診断し、GitHub CLIを使ってエラーログを抽出し、修正計画を生成します。ユーザーがCIステータスの確認、CI問題の取得、テスト失敗のトリアージ、またはPRビルド障害の調査を依頼した場合に使用します。
developmenttestingdevops
fix-errors
warpdotdev
Warp Rustコードベースのコンパイルエラー、リンティング問題、テスト失敗を修正します。presubmitチェック、WASM固有のエラー、特定のテストの実行をカバーします。ユーザーがビルドエラー、clippyやfmtの失敗、テスト失敗に遭遇した場合、またはPR前にpresubmitを実行・解釈する必要がある場合に使用します。
developmentcode-reviewtesting
implement-specs
warpdotdev
承認されたPRODUCT.mdとTECH.mdの機能を実装し、実装の進化に合わせて仕様とコードを同じPR内で整合させます。プロダクトと技術仕様が承認され、次のステップが機能の構築である場合に使用します。
developmentcode-reviewapi