菜单
Orchestrator that evaluates a built skill by running agent-audit, agent-quality, or both. Asks the user which agents to run and which mode (flash or comprehensive). Use when the user says "evaluate this agent", "run evaluation on", or invokes /agent-evaluate after agent-build completes.
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
Senior product designer that audits UI screens and PRDs for usability flaws, missing error states, edge-case gaps, and accessibility issues. Use when the user says "review this design", "audit this UI", "critique this flow", or invokes /design-review with a screenshot and a PRD.
Fifth subagent in the agent-audit pipeline. Reads evals-[n].json and grading.json from the run dir, aggregates timing across all evals, computes pass rate and token stats, and writes timing.json and benchmark.json. Handles missing or partial data gracefully — writes what it can and flags gaps. Use when agent-audit hands off "compute the benchmark".
Second subagent in the agent-audit pipeline. Grades every assertion in evals-[n].json against actual_output using LLM judgment for semantic checks and tool calls for mechanical checks. Writes grading.json with per-assertion pass/fail and evidence, and updates evals-[n].json with a verdict per case. Use when agent-audit hands off "grade the evals".
Third subagent in the agent-audit pipeline (comprehensive mode only). Checks the environment for agentlinter and agnix, installs missing tools via npm if possible, runs each linter independently, then runs an LLM safety check against the audit registry. Each check is independent — a missing or failing tool is flagged and skipped, never blocks the others. Writes audit-[n].json to the run dir. Use when agent-audit hands off "run lint and safety check".
Fourth subagent in the agent-audit pipeline (comprehensive mode only). Measures the target skill's description trigger rate, iterates to improve it, and writes the winning description back to SKILL.md. Optimised for token efficiency: train iterations use 1 run per query; validation and baseline use 3. Stops early at 90% train pass rate. Appends optimiser results to audit-[n].json. Use when agent-audit hands off "run the description optimiser".
Orchestrator for the agent-audit pipeline. Prepares the run directory, asks the user which checks to run via a labelled checklist, dispatches specialist subagents in the correct dependency order, and produces a final report. Use when the user says "audit this skill", "run agent-audit", "check this agent", or agent-evaluate hands off "run audit".
| name | agent-evaluate |
| description | Orchestrator that evaluates a built skill by running agent-audit, agent-quality, or both. Asks the user which agents to run and which mode (flash or comprehensive). Use when the user says "evaluate this agent", "run evaluation on", or invokes /agent-evaluate after agent-build completes. |
| allowedTools | ["Read","Skill","AskUserQuestion","TodoWrite"] |
Invoke: /agent-evaluate <skill-name> — pass the skill name matching .claude/skills/<name>/.
/agent-evaluateagent-build emits "Ready for /agent-evaluate"| Name | Format | Source |
|---|---|---|
| skill_name | directory name under .claude/skills/ | args or prior agent-build context |
| Name | Format | Destination |
|---|---|---|
| evaluation_report | markdown grouped by agent (audit / quality); findings with severity and fix | shown inline |
| summary | ≤5-bullet pass/fail summary across selected agents | shown inline (after report) |
refs/sub-agent-contracts.md. Knows what flash and comprehensive mode mean for each sub-agent. Knows how to aggregate conflicting severity calls across agents.SKILL.md is missing, names the missing file and stops. Does not guess or infer a path.Emit Step X/5 — <title> at the start of each step, unconditionally.
Step 1/5 — Resolve skill path
Read skill_name from args. If absent, check prior context for a skill name emitted by agent-build. If still absent, ask once: "Which skill should I evaluate? Pass the name matching .claude/skills/<name>/." Verify .claude/skills/<skill_name>/SKILL.md exists. If missing, emit: Cannot evaluate: .claude/skills/<skill_name>/SKILL.md not found. End the run. Produce skill_path = .claude/skills/<skill_name>/.
Step 2/5 — Select agents Ask the user:
Which agents should I run?
- Both — audit then quality
- Audit only (recommended minimum) — structural completeness check
- Quality only — content depth and language check
Wait for the answer. Store agents as [audit, quality], [audit], or [quality]. Produce agents.
Step 3/5 — Select mode Ask the user:
Which mode?
- Flash — top 3 findings per agent, only P0 issues
- Comprehensive — exhaustive findings, ranked by severity
Wait for the answer. Store mode as flash or comprehensive. Produce mode.
Step 4/5 — Run selected agents
Invoke the selected agents using the Agent tool. If both agents are selected, send two Agent tool calls in a single message so they run in parallel — do not wait for one to finish before starting the other. If only one agent is selected, invoke it alone. Pass skill_path and mode as args to each. Follow refs/sub-agent-contracts.md for the exact invocation interface. Collect each agent's output as audit_output and/or quality_output once both complete.
Step 5/5 — Aggregate and emit
Render one report section per agent run. Label each section clearly (## Audit findings / ## Quality findings). In flash mode, include only the top 3 findings per section sorted by severity. In comprehensive mode, include all findings sorted by severity then file order. After both sections, render ## Summary — up to 5 bullets stating overall pass/fail per agent and the single highest-priority fix from each. End the run.
refs/sub-agent-contracts.md — interfaces, args, and expected output shape for agent-audit and agent-quality