メニュー
Loaded at session start. Establishes awareness of craft-skills so the assistant invokes relevant skills before acting. Not invoked directly.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | bootstrap |
| description | Loaded at session start. Establishes awareness of craft-skills so the assistant invokes relevant skills before acting. Not invoked directly. |
You have access to the craft-skills plugin — a set of structured development skills for any project.
If there is even a 1% chance that a craft-skill applies to what you are about to do, you MUST invoke it via the Skill tool BEFORE taking action. Skills override your default behavior with battle-tested workflows.
| Skill | Trigger Conditions |
|---|---|
craft-skills:debug | Bug report, error investigation, failing tests, unexpected behavior. Invoke BEFORE attempting any fix. |
craft-skills:reflect | User asks to audit, improve, or maintain Claude configs/skills. Also: periodic maintenance. |
craft-skills:simplify | After implementation is done, review changed code for reuse/quality. |
| Skill | Trigger Conditions |
|---|---|
craft-skills:craft | Default craft pipeline (Claude only, no external deps). Use when requirements need deep exploration and you want the simplest, most reliable path. |
craft-skills:craft-ace | Craft with the local LLM (Qwen3.6) implementing and reviewing, Claude orchestrating. Use for cost savings when LM Studio is running. |
craft-skills:implement | New feature with clear requirements. User says "implement", provides a prompt file number, or gives clear feature description. |
craft-skills:finalize | Plan already exists (from brainstorming or architect). User says "finalize", "execute the plan", or references an existing plan file. |
craft-skills:architect | User wants only planning, no implementation. User says "plan", "architect", or "how should we approach". |
craft-skills:develop | Approved plan exists, ready to execute. User says "develop", "build it", or references a plan to execute. |
craft-skills:browser-test | Feature is built, needs browser testing. User says "test in browser", "check the UI", or testing is the next pipeline step. |
craft-skills:finish | Implementation + verification done; integrate the work. User says "finish", "commit", "open a PR", "merge", or "ship it". The final pipeline step. |
| Skill | Usage |
|---|---|
craft-skills:graph-explore | Reference for graph MCP tool queries. Calling skills run graph tools directly via ToolSearch — no agents. |
craft-skills:llm-review | Reference for local LLM scripts. Calling skills run bash commands directly — no agents. |
How these work: Graph MCP tools and LLM bash scripts run directly in the main conversation. Do NOT dispatch dedicated agents for graph or LLM — agents cannot reliably run bash scripts and add unnecessary indirection. Each calling skill contains the exact bash commands and graph tool sequences inline.
When multiple skills could apply:
For atomic operations, route to craft skills — do NOT fall through to another plugin's standalone skills (different conventions, different dirs):
| Request | Use |
|---|---|
| "write a plan" / "plan this" | craft-skills:architect — plans live in .claude/plans/, NOT docs/superpowers/ |
| "brainstorm" / "explore this idea" | craft-skills:craft Phase 1 — specs live in .claude/plans/specs/ |
| "finish" / "commit" / "open a PR" / "ship it" | craft-skills:finish |
If a .craft-profile, .shared-state.md, or .craft-progress.md exists (a craft run is in progress), craft skills take precedence over any other plugin's plan/spec/execute/finish skills. Use craft's dirs (.claude/plans/), not docs/superpowers/. Do NOT invoke another plugin's test-driven-development or using-git-worktrees inside a craft run — craft verifies via lint/tsc/build and works in the current checkout.
Before any implementation work, check:
Do NOT skip skills because:
| Excuse | Reality |
|---|---|
| "This is too simple" | Simple tasks have the most unexamined assumptions |
| "I already know the answer" | Skills enforce process, not just answers |
| "It would be faster without the skill" | Skipping skills causes rework that wastes more time |
| "The user didn't ask for it" | Skills are invoked automatically when relevant |
| "I just need to make a quick fix" | debug skill exists precisely for this — investigate first |
| "I'll just explore first" | Exploring IS the first step of the relevant skill |
finishWhen no skill applies, proceed normally following the project's CLAUDE.md conventions.
Use when the user needs to plan and architect a new feature before implementation begins. This includes analyzing requirements, asking clarifying questions, and creating a detailed implementation plan aligned with the project's architecture. Invoke BEFORE any coding work starts.
Use when implementation + verification are complete and the work needs to be integrated — commit, open a PR, merge, or clean up the branch. The final step of the craft pipeline (after develop/test/report). Works in the current checkout; no worktree assumption.
Use when starting a new feature that needs thorough design exploration. This is the most comprehensive workflow: collaborative brainstorming with the user, then planning, then parallel agent development, then browser testing. Use for complex features, unclear requirements, or when the user wants to explore approaches before committing.
Post-develop visual regression + design-contract audit. Starts the dev server, captures screenshots of changed UI routes, and dispatches a Haiku-vision agent to compare the result against .claude/aesthetic-direction.md and the feature's ux-brief.md. Flags regressions before browser-test runs. Escalates to Sonnet or Opus only when Haiku flags issues.
Use when an approved implementation plan exists and is ready to be executed. Initializes shared state, dispatches parallel developer agents, performs integration review, and runs full verification (lint, tsc, format, build).
Generate or refresh .claude/reuse-index.md — a one-page inventory of the project's shared utilities, enums, hooks, and components that downstream implementer agents must consult before writing new code. Project-agnostic scan: no framework assumptions baked in. Invoke once per project (or when the shared surface changes meaningfully).