菜单
Enforces subagent frontmatter and system-prompt conventions when a `<name>.md` agent file is opened for authoring or review — plugin-scope ignored fields, description quality, tool-restriction shape, skills preload, and the required boundaries section.
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。
基于 SOC 职业分类
| name | agent-authoring |
| description | Enforces subagent frontmatter and system-prompt conventions when a `<name>.md` agent file is opened for authoring or review — plugin-scope ignored fields, description quality, tool-restriction shape, skills preload, and the required boundaries section. |
| user-invocable | false |
| paths | ["**/agents/**/*.md","**/.claude/agents/**/*.md"] |
Apply this to the file you just opened. The two mistakes that matter most: authoring hooks/mcpServers/permissionMode into an agent that only ever ships as a plugin (silently ignored at load), and a description that narrates the agent's steps instead of stating when to delegate to it.
name and description present — the only required fields.hooks, mcpServers, and permissionMode in its frontmatter are dead. Claude Code ignores all three when the agent file loads from <plugin>/agents/. Only flag them if (a) they're load-bearing for the agent's intended behavior and (b) plugin scope is the only delivery path — then propose moving the hook to <plugin>/hooks/hooks.json (matcher scoped to agent_type) or the MCP server to the plugin manifest's mcpServers block. Otherwise the fields are harmless portability leftovers for project/user/CLI scope; don't strip them reflexively.tools: matches what the system prompt actually calls. Prefer an explicit allowlist or disallowedTools over omitting tools: — inheriting everything makes a plugin agent unauditable.skills: preloads only always-needed content. Each entry injects full skill content at every startup — expensive. Conditional knowledge belongs in description- or path-triggered skills the agent discovers on demand, not here.haiku for read-only search, sonnet for mechanistic work, opus/inherit for reasoning-heavy tasks.isolation: worktree set for agents that make destructive or exploratory changes to the working tree.memory: scope, if set, has consult-before/update-after instructions in the body — otherwise the memory directory exists but goes unused.skills:. An agent that will always need hook-scripts or skill-scripts context should list it in skills: rather than relying on the agent discovering it via path-trigger mid-task — startup cost is worth it for skills the agent needs on turn one.hooks/mcpServers/permissionMode frontmatter. Design the agent's behavior assuming these three are inert at plugin scope. If a capability genuinely requires one of them, it does not belong in a plugin-shipped agent — ship it at project/user scope instead, or move the mechanism to the plugin's own hooks.json / plugin.json.tools: Bash unscoped when the agent only runs git and npm — tighten to Bash(git *), Bash(npm *).disable-model-invocation or paths: in agent frontmatter — those are skill fields, not agent fields.${CLAUDE_PLUGIN_ROOT}/skills/anthropic-docs/references/subagents.md — every frontmatter field, plugin-scope restrictions, model resolution order, tool-restriction syntax, memory scopes, hooks/mcpServers inline schemas.Reviewing whether the system prompt's language will actually land — imperative force, urgency calibration, structure? Invoke the persuasion skill.
Runs a dogfood cycle against plugin-bot — tasks the plugin-engineer agent with building a plugin capability inside the plugins/dogfood sandbox, validates and reloads it, evaluates the result (skill-creator evals for skills, fixtures/BATS for hooks), and harvests rough edges in plugin-bot's own skills and agent as improvement notes. Use after changing plugin-bot to exercise it end to end.
Distilled, verification-stamped reference for the official Claude Code platform contract — plugins, skills, subagents, hooks, MCP, tools, env vars and channels. Use when authoring or auditing any Claude Code plugin component, looking up a frontmatter field, hook event schema, manifest rule, permission-rule format or tool name, or whenever a platform claim needs verifying instead of recalling from memory.
Enforces the house hook-script contract when a `hooks/**/*.sh`, `hooks/**/*.bash`, or `hooks/hooks.json` file is opened for authoring or review — envelope/exit-code/JSON discipline, subdirectory-per-event layout, lib helpers, fixtures, and BATS coverage.
Use when authoring, auditing, or registering a Claude Code plugin monitor — a monitors/monitors.json entry or the background watch script it runs. Ships the typed TypeScript poll-monitor harness (self-scheduling loop, stable-streak debounce, notify-once dedup, --once mode) plus the registration rules and notification-copy doctrine; auto-loads as the enforcer when a monitors/ file is opened.
Use when writing or reviewing language meant to shape an agent's behavior — a hook's `additionalContext` payload, a SKILL.md body or description, an agent system prompt, or an enforcer checklist. Covers choosing the imperative force, structuring with XML tags, matching the urgency tier to the moment (the EXTREMELY_IMPORTANT / IMPORTANT / reminder / acceptable-anti-pattern gradient), the persuasion-principles foundation (Cialdini + Meincke), why XML beats markdown for hook injections, and concrete templates for SessionStart, PostToolUse, and PreToolUse nudges.
Enforces manifest well-formedness when `.claude-plugin/plugin.json`, `.claude-plugin/marketplace.json`, `.mcp.json`, or `.lsp.json` is opened for authoring or review — path-substitution rules, the hooks.json separation, and the house version policy.