菜单
Scaffold a new Claude Code plugin and wire its components. Use when the user asks to create or scaffold a plugin, package skills, agents, or hooks as a plugin, or make a workflow installable/shareable. Not for auditing an existing plugin — route that to validate-plugin.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
User-invoked pass that re-derives a project's Claude-native instruction layout — a surface AGENTS.md plus path-scoped .claude/rules/ — as a projection of the installed skill set and the actual codebase. Enumerates every available skill, maps which applies to which path (wiring unwired-but-needed ones, dropping inapplicable ones), grounds every claim against the code, discloses only non-code-recoverable info per path, and drops the inert references/ layer. Read-only toward user-owned notes (roadmaps, scratchpads) — it reports promotion candidates, never rewrites them. Invoke as /meta-dev:revise-agents-md.
Bootstrap a project's instruction files from scratch — a surface AGENTS.md plus path-scoped .claude/rules/ derived from the installed skills and the actual code. Use when the user asks to set up CLAUDE.md/AGENTS.md, bootstrap instruction files or project memory, or initialize the layout for a repo that has none. Not for editing an existing layout's content (update-instructions), tightening one already in place (improve-instructions), or re-deriving the whole layout of a repo that already has one (revise-agents-md).
Raise the quality of a project's instruction files (CLAUDE.md, AGENTS.md, references) WITHOUT changing what they commit to. Use when the user asks to clean up, debloat, or tighten CLAUDE.md or AGENTS.md, audit instruction files, or says their constitution is bloated, messy, or hard to read. Fact or guard changes belong to update-instructions; first-time bootstrap belongs to setup-instructions.
Raise an existing skill's quality against a measurable criterion WITHOUT changing its intended behavior. Use when the user asks to debloat, tighten, shorten, or clean up a skill, fix its formatting or wording, rewrite a description already diagnosed as under-triggering (diagnosing why belongs to review-skill), or run the automated description-optimization loop. Behavior changes belong to update-skill; skills over their size budget belong to restructure-skill.
Reorganize skill architecture — split one skill into several, merge overlapping skills, or move content between SKILL.md, references, and scripts. Use when a skill exceeds its size budgets, when skills overlap or their descriptions collide, when content sits at the wrong loading level, or when the user asks to split, merge, rename, or reorganize skills. Content edits within one healthy skill belong to update-skill or improve-skill.
Read-only audit of a project's instruction files (AGENTS.md, CLAUDE.md, .claude/rules) — never edits them. Use to review, check, audit, or diagnose instruction files, verify every documented claim against the actual codebase, find stale or wrong facts, spot cross-file contradictions or bloat, or get a quality report before deciding what to change. Not for applying fixes (update-instructions / improve-instructions) or first-time bootstrap (setup-instructions).
基于 SOC 职业分类
| name | scaffold-plugin |
| description | Scaffold a new Claude Code plugin and wire its components. Use when the user asks to create or scaffold a plugin, package skills, agents, or hooks as a plugin, or make a workflow installable/shareable. Not for auditing an existing plugin — route that to validate-plugin. |
| model | sonnet |
Scaffold a new plugin directory, wire its components, and hand off for validation. Read ${CLAUDE_SKILL_DIR}/../../standards/core.md and ${CLAUDE_SKILL_DIR}/../../standards/plugin.md before drafting — directory layout, manifest fields, and distribution mechanics live there and only there.
This is create-class work: per core.md's autonomy split, ask during intake only, then proceed hands-free.
~/.claude/skills/ carrying a .claude-plugin/plugin.json, loading as <name>@skills-dir — for dev/personal use with no install step; or standalone repo + marketplace — for shareable installs across machines/users. The choice changes which later steps run (marketplace.json only applies to the second).Completion criterion: name, component list, and distribution intent all confirmed by the user.
claude plugin init <name> --with <skills|agents|hooks...> (writes to ~/.claude/skills/<name>/), then adjust the generated manifest. Hand-author only for standalone-repo intent, where init's target directory is wrong: create .claude-plugin/plugin.json with at minimum name. Set version if the user wants updates pinned to explicit bumps (per plugin.md: omitting it under git distribution means every commit counts as a new version)..claude-plugin/ — skills/, agents/, hooks/ as needed. Do not create a commands/ directory; it is legacy and this skill authors skills only.CLAUDE.md for context — plugin.md states it is not loaded; any knowledge the plugin needs must ship through skills, agents, or hooks content instead.Completion criterion: .claude-plugin/plugin.json exists with valid name, and every requested component directory exists at plugin root with nothing but the manifest under .claude-plugin/.
skills/<name>/SKILL.md with correct frontmatter (name matches folder, description is triggers-only per core.md, model set deliberately).agents/<name>.md, and check the plugin-agent field limits in plugin.md/agent conventions before assuming full parity with standalone agents.hooks/hooks.json, referencing any bundled scripts via ${CLAUDE_PLUGIN_ROOT} rather than a hardcoded path — a plugin cannot reach files outside its own directory, and hardcoding breaks the moment the plugin is installed to a different cache location.Completion criterion: every component named in Step 1 exists in its correct root-level directory, and all intra-plugin paths use ${CLAUDE_PLUGIN_ROOT}.
Skip this step entirely for skills-dir plugins — they load in place with no marketplace involved.
.claude-plugin/marketplace.json with name, owner, and plugins: [{name, source: "."}].claude plugin marketplace add <path> then claude plugin install <plugin>@<marketplace>.Completion criterion: marketplace.json validates and the plugin installs successfully from it.
validate-plugin against the scaffolded directory for the authoritative structural/manifest check — this skill builds the plugin, it does not certify it; a second, focused pass catches what drafting attention misses./reload-plugins for skills-dir, or the marketplace install command from Step 4), and any findings validate-plugin returned.Completion criterion: validate-plugin has run against the final layout and the user has the install/reload instructions in hand.