跳到正文
cccskills
免费在 GitHub 公开

brainstorming

Use before implementing any feature, behavior change, or refactor - settles requirements and design, then gets the design independently reviewed and approved by the user before code is written

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(8)

  • SKILL.md6.1 KB
  • scripts/frame-template.html7.9 KB
  • scripts/helper.js5.5 KB
  • scripts/server.cjs25.1 KB
  • scripts/start-server.sh6.7 KB
  • scripts/stop-server.sh3.2 KB
  • spec-document-reviewer-prompt.md584 B
  • visual-companion.md13.1 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Brainstorming Ideas Into Designs

Read the relevant existing flow and the user's constraints. Identify the intended behavior, scope, and observable success criteria. Do not reopen decisions already given.

  • For a bounded change, a short design in chat is enough: approach, files touched, and tests.
  • For exploratory work, state what the probe can establish and report its limitations. Do not turn a disposable experiment into a product change outside the authorized scope.
  • For architectural work, compare the meaningful alternatives and document interfaces, data flow, failure behavior, compatibility, and verification. Split only where the work has independent deliverables.

Bundle related clarification questions into one message. Resolve routine implementation choices from the repository and conversation. When a consequential product decision or authorization is missing, ask and continue only independent work while waiting. Do not interpret elapsed time as approval.

Write a spec when it helps review or preserves decisions across sessions, normally at docs/superpowers/specs/YYYY-MM-DD-topic-design.md. Check it for ambiguity, contradictions, and scope gaps before the design review. Respect ignored/local-only documentation and do not force a commit.

Design Gate

When the design is complete, complete these steps in order before writing any implementation code. The gate applies to bounded and architectural designs alike; only the length of the design scales.

  1. Create a task branch first. Review reservations are per branch and are refused on main/master, so create the task branch before reserving the design review.
  2. Reserve the review. Run node .claude/hooks/review-budget.cjs begin --phase design --roles document-reviewer (.codex in a Codex-only project; in an environment where the review-budget hook is not installed or not trusted (in this distribution, Cursor for example), npx --no ai-dev-helm review-budget begin with the same arguments (where the CLI is not installed, npx -y @crearize/ai-dev-helm@<version in .ai-dev-helm.json> review-budget begin); see documents/development/harness-runtime.md). Add --limit 3 to this first reservation only when the design is known to be large (quality-policy §5.5). Place the returned marker line at the top of the reviewer prompt.
  3. Request the design review. Dispatch one reviewer named document-reviewer (Codex: helm-doc-reviewer) with the design-review model from harness-runtime.md. Give it spec-document-reviewer-prompt.md, the design or spec path, the user's requirements, and the relevant existing code. Fix confirmed findings. Another round is allowed only within the reserved limit and when the review raised a High finding; otherwise report the remaining concern in step 3 rather than reserving again.
  4. Ask the user to review the design. Send one message with the design, the review findings and how each was resolved, and the decisions that need approval. Stop there. Start implementation only after the user explicitly approves. If the user requests changes, revise and present the design again; ask the user before any further review round. If the user approves exceeding the review limit, run extend yourself with the same review-budget script used in step 1 (.claude/hooks/…, .codex/hooks/…, or npx --no ai-dev-helm review-budget (where the CLI is not installed, npx -y @crearize/ai-dev-helm@<version in .ai-dev-helm.json> review-budget)): extend --phase design --rounds <1-3> --reason "<the user's approval, quoted>"; never ask the user to run commands, inspect state, or reset state.

Exemption: a typo fix that involves no design choice, or an obviously correct change of a few lines. A fix within an already approved design (such as a response to a quality-check finding) that does not change that design needs no new Design Gate. State what you will change before making an exempt change.

None of these replaces the gate: your own self-check, a request to implement given before the design existed, a plan review, or the final quality-check. If the reservation is denied, follow the hook's message; ask the owner for approval only when a review limit is reached, then run extend yourself.

Once the user approves, do not ask again. The approval covers the spec update, the plan and its review, implementation, tests, documentation, quality-check, commits, feature-branch pushes, the PR and the merge when no deviation from the design was recorded (with deviations, a final confirmation before merging) (documents/development/development-policy.md §1.0 "承認後の進め方").

After Approval

Update the spec, if any, with the approved decisions. When implementation has to depart from the design, choose what best fits the design's intent and approved decisions, keep going, and record the deviation (what -> what you did / why / impact) in a final "## 実装時の差異" section of the spec (or the plan, or a working note) before quality-check creates its flag. A deviation found after the flag is recorded there and in the PR body too; delete the flag, then ask about merging with it; after the OK, create the flag on that HEAD (quality-check Step 6) when the record is the only change since the check (§1.0 "承認後の進め方" 1 and 3). Do not report or consult mid-way; the final message reports only the deviations. Stop mid-way only for the exceptions X1-X8 in that section; X4 (the design cannot reach its goal) asks once for the deviation design and, if no design-review round is left, one more round. Use writing-plans when sequencing or handoff benefits from a plan; otherwise continue with implementation and appropriate tests. Preserve the independent final quality-check.

For decisions that benefit from browser mockups, use the optional visual companion. Follow the user's existing preference and the guide's browser launch instructions.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

バックエンド実装時に使用。DRY原則遵守。コーディング規約準拠。

中文简介正在准备中。当前显示原文说明。

Crearize/ai-dev-helm52026年10月12日 更新

作業開始時に使用。mainブランチでの作業禁止。Issue先行作成必須。

中文简介正在准备中。当前显示原文说明。

Crearize/ai-dev-helm52026年10月12日 更新

UI実装後の検証時に使用。agent-browser CLIでブラウザ上の動作を手動検証する。「UIを確認」「画面テスト」と言われたら使用(プロジェクトの E2E スイート実行は quality-check Step 5(推奨度・範囲で自動実施または確認)/ server-startup が担当)。

中文简介正在准备中。当前显示原文说明。

Crearize/ai-dev-helm52026年10月12日 更新

DBマイグレーション作成時に使用。バージョン番号競合防止。mainブランチ確認必須。

中文简介正在准备中。当前显示原文说明。

Crearize/ai-dev-helm52026年10月12日 更新

Use when independent tasks benefit from parallel work without shared state or sequential dependencies

中文简介正在准备中。当前显示原文说明。

Crearize/ai-dev-helm52026年10月12日 更新

Use when continuing implementation from an existing plan

中文简介正在准备中。当前显示原文说明。

Crearize/ai-dev-helm52026年10月12日 更新

查看 Crearize 的全部技能

报告此技能的问题