菜单
Review an existing agent against best practices and report findings — a read-only audit of CLAUDE.md, skills, composition integrity, and Trinity readiness. Produces a prioritized findings report and hands off to /adjust-agent or /adjust-playbook to apply fixes. Makes no changes itself.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Abilityai/abilities --skill review-agent命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Generate an agent-specific `/update-dashboard` skill that keeps `dashboard.yaml` current for Trinity. Analyzes the agent's purpose and data sources, proposes metrics, gets user approval, then scaffolds a schedulable skill.
Modify an existing playbook based on conversation context or explicit instructions. Use when user wants to update, fix, extend, or refine a playbook they already have.
Create a new skill or playbook. Guides through requirements gathering and generates the appropriate template based on complexity.
正在显示 SKILL.md
| name | review-agent |
| description | Review an existing agent against best practices and report findings — a read-only audit of CLAUDE.md, skills, composition integrity, and Trinity readiness. Produces a prioritized findings report and hands off to /adjust-agent or /adjust-playbook to apply fixes. Makes no changes itself. |
| argument-hint | [path to agent] [focus area] |
| disable-model-invocation | false |
| user-invocable | true |
| allowed-tools | Read, Glob, Grep, Bash, Skill |
| metadata | {"version":"1.6","created":"2026-06-14T00:00:00.000Z","updated":"2026-08-18T00:00:00.000Z","author":"Ability.ai","changelog":["1.6: Audit checklist adds the .mcp.json.template URL rule from Trinity v0.9.0 — an http/sse server url must resolve to a public address (loopback/private/link-local/CGNAT 100.64/10 refused, trinity-enterprise#394)","1.5: Playbook-call checks — schedule messages must be one-line `/skill [args]` calls (no prose briefs), and inter-agent hand-offs in CLAUDE.md/skills must be playbook calls, not prose delegation (fleet convention protocols/playbook-call.md, operator direction 2026-08-16)","1.4: Audit checklist matches the current platform contract — schedule entries key on `name` (there is no `id` field), template.yaml must declare credentials: + credential_setup: (gate T-015, ent#128), .mcp.json.template must keep ${VAR} inside env blocks with an allowlisted command and no hand-written trinity entry, and .gitignore must exclude .claude/settings.json (trinity#2036) and .trinity/*","1.3: Trinity-readiness check (2h) now audits repository deployability — remote present, tree clean, branch pushed, plus the instance GitHub token note for private repos — because Trinity deploys an agent by cloning its repo; a repo-less agent is reported as deploy-by-upload-only (no reproducible source) with the one-command fix","1.2: Audit Request Dispatch (new check 2k) — CLAUDE.md carries the SOP routing table: request-phrased rows, every user-invocable skill covered or deliberately internal, routes resolve, playbook-gap fallback with operator-queue escalation and /agent-dev:create-playbook pointer; scorecard + checklist rows added","1.1: Audit Documentation Coherence — README/ARCHITECTURE/TARGET-ARCHITECTURE present, current/target split intact, docs match skills and subagents","1.0: Initial version — read-only agent audit that hands off to /adjust-agent"]} |
ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of
metadata.changelogabove — e.g.review-agent vX.Y — recent: <summary>. Then proceed.
Audit an existing agent against best practices and produce a findings report. Read-only by construction — this skill has no Write/Edit tools and changes nothing. It is the detect-and-report half of the pair; /adjust-agent (agent-level) and /agent-dev:adjust-playbook (skill-level) are the apply half.
This skill is the single source of truth for the agent audit rubric. /adjust-agent invokes this skill to get its findings rather than duplicating the checks.
/adjust-agent or /agent-dev:adjust-playbookIf $0 (path) provided, use it. Otherwise assume the current working directory.
Verify it's an agent:
ls CLAUDE.md 2>/dev/null && ls -d .claude/skills 2>/dev/null
If CLAUDE.md doesn't exist, ask for the correct path.
Read all agent artifacts (a missing one is itself a finding):
CLAUDE.mdREADME.md, ARCHITECTURE.md, TARGET-ARCHITECTURE.md (a missing one is a Documentation Coherence finding)template.yaml.env.example.gitignore.mcp.json.template or .mcp.jsonSKILL.md files: find .claude/skills -name "SKILL.md" 2>/dev/nullls .claude/agents/*.md 2>/dev/nullls -laIf $ARGUMENTS names a focus area (e.g. "composition", "schedules", "trinity"), audit only that area; otherwise run the full audit.
Evaluate each area and assign a status:
Detect and classify only — do not draft fixes (that's /adjust-agent's job).
## Identity section with the agent name in bold## Core Capabilities (or equivalent) section/skill-name).claude/skills/ is listed, and every listed skill exists## Artifact Dependency Graph sectionartifacts: block with mode (prescriptive/descriptive) and direction (source/target)sources declared for target artifactssync_skills: mapping skills to source→target edgesschedules: block in template.yaml (the design source of truth)name, cron, message (optionally timezone, purpose, enabled) — name is the identity key Trinity dedups on; there is no id field, and a stray id: is silently ignoredenabled: false by default; a ## Recommended Schedules table in CLAUDE.md renders the blockmessage is a one-line playbook call — /skill [args], ≤120 chars, no prose briefing or restated policy — and names a skill that exists (fleet convention protocols/playbook-call.md / schedule-contract.md; a prose message is unversioned process the platform cannot measure)## Guidelines section with 2-4 domain-specific, actionable rules (not generic advice)For each SKILL.md:
name, description, allowed-tools, user-invocableallowed-tools is comma-separated (not a YAML array)metadata.version presentA cross-skill graph check — the part a per-file audit can't do. First build the call graph: which skills invoke which.
# Candidate invocations across all skills (review each by hand — see the two traps below)
grep -rnE "Invoke \`/|^Skill: " .claude/skills/*/SKILL.md 2>/dev/null
Two false-positive traps — exclude these before reporting (see Composing skills):
add-*/scaffold skill that cps skill files into a target agent is installing deliverables, not composing — not a finding.Invoke `/x` inside a code fence, in a skill whose job is to generate skill text, is documentation, not a call — not a finding.Real findings:
Invoke `/child` but lacks Skill in allowed-tools (the call can't run).automation: autonomous skill that invokes a gated/manual child (transitive autonomy — the heartbeat would hang)./child-vN while an unversioned /child exists, so it won't receive the child's fixes.template.yaml with name, display_name, description, avatar_prompt, and a credentials: block naming every ${VAR} used in .mcp.json.template (gate T-015), each enriched by a credential_setup: entry (ent#128).env.example documenting required variables.mcp.json.template declaring the agent's own MCP servers, with ${VAR} placeholders inside env blocks only (a placeholder in command/url/args makes Trinity withhold the whole server at startup) an allowlisted literal command, and — for http/sse servers — a url that resolves to a public address (loopback/private/link-local/CGNAT 100.64/10 are refused with 400, trinity-enterprise#394) — never a hand-written trinity entry, which the platform overwrites.gitignore excluding .env, .env.*, .mcp.json, credentials.json, *.pem, *.key, .claude/projects/, .claude/todos/, .claude/plugins/, .claude/settings.json (trinity#2036), and .trinity/* with the authored hooks negatedgit remote exists (git remote get-url origin)gh repo create <name> --private --source=. --pushThe agent should keep an honest current→target picture and a human-facing overview.
README.md present — human-facing capabilities overview, and its skill/capability list matches the actual skillsARCHITECTURE.md present — describes current state; components it names (skills, subagents, data, schedules) actually exist on disk, and nothing real is undocumentedTARGET-ARCHITECTURE.md present — describes target state; the current/target split is intact (no item described as shipped/live in ARCHITECTURE.md is still listed as "planned" in the target doc)mode/direction (README + ARCHITECTURE descriptive targets; TARGET-ARCHITECTURE prescriptive source)/reconcile-docs skill exists to maintain this coherence (and ideally is scheduled)This is the detect counterpart to a generated agent's own /reconcile-docs — the agent self-checks; this audit verifies the machinery is present and the docs aren't already drifting.
The agent's SOP for incoming requests — from the user, other agents, or the operator queue.
## Request Dispatch section in CLAUDE.md with a request-routing table.claude/skills/ has a dispatch row, or is deliberately internal (e.g. a vendored helper composed by another skill)playbook-gap-<slug> operator-queue item when headless on Trinity — with a pointer to /agent-dev:create-playbook/playbook [args]) and expect the same in return (protocols/playbook-call.md)Produce a findings report — no edits, no AskUserQuestion to apply anything. Order findings highest- to lowest-impact: structural gaps → completeness → quality → polish.
## Agent Review: [name]
**Path**: [path]
**Skills**: [count]
**Overall**: [healthy / needs work / incomplete]
### Findings
| # | Area | Status | Finding | Severity |
|---|------|--------|---------|----------|
| 1 | Composition | BROKEN | pipeline-tick invokes /foo, which doesn't exist | High |
| 2 | Identity | IMPROVE | Purpose statement missing | Medium |
| … | | | | |
### Scorecard
| Area | Status |
|------|--------|
| Identity | [PASS/IMPROVE/MISSING] |
| Capabilities | … |
| Dependency Graph | … |
| Schedules | … |
| Guidelines | … |
| Skill Quality | … |
| Composition Integrity | … |
| Documentation Coherence | … |
| Request Dispatch | … |
| Trinity Ready | … |
| Hygiene | … |
### Apply fixes
This review changed nothing. To act on it:
- **Agent-level** (identity, graph, schedules, Trinity files): `/adjust-agent [path]`
- **Skill-level** (a specific SKILL.md — steps, permissions, composition): `/agent-dev:adjust-playbook [skill-name]`
| # | Area | Check | Priority |
|---|---|---|---|
| 1 | Identity | Named, purposeful, audience-aware | High |
| 2 | Capabilities | Listed, linked to skills, when-to-use | High |
| 3 | Dependency Graph | Artifacts declared, directions set, skills mapped | Medium |
| 4 | Schedules | Automatable skills scheduled with cadences; every message is a one-line playbook call | Medium |
| 5 | Guidelines | 2-4 domain-specific actionable rules | Medium |
| 6 | Skill Docs | CLAUDE.md ↔ .claude/skills/ in sync | High |
| 7 | Skill Quality | Frontmatter valid, permissions minimal, steps clear | High |
| 8 | Composition Integrity | Invocations resolve, no cycles, autonomy transitive, Skill tool present | High |
| 9 | Documentation Coherence | README/ARCHITECTURE/TARGET-ARCHITECTURE present, current/target split intact, docs match skills & subagents, /reconcile-docs exists | Medium |
| 10 | Request Dispatch | SOP routing table in CLAUDE.md — request-phrased rows, routes resolve, all user-invocable skills covered, playbook-gap fallback | Medium |
| 11 | Trinity Ready | template.yaml, .env.example, .gitignore, .mcp template | Low |
| 12 | Hygiene | Git init, no secrets, naming convention | Low |
| Skill | Purpose |
|---|---|
| /adjust-agent | Apply the fixes this review surfaces (agent-level) |
| /agent-dev:adjust-playbook | Apply fixes to a single skill |
| /create-agent | Scaffold a new agent from scratch |