菜单
Inspect the current AxonFlow repository and runtime, identify evidence-backed product gaps, and turn the highest-value gap into one bounded implementation task. Use for product health reviews, self-optimization runs, backlog proposals, and deciding the next improvement before coding.
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。
Run automated AxonFlow acceptance tests, operate the local UI in a real headless browser, capture visible evidence, and record every requirement as passed or failed. Use after coding changes, for workflow UI regression checks, and whenever HTTP-only checks cannot verify rendered or interactive behavior.
Review an AxonFlow product task before coding, validate evidence, value, scope, safety, and testability, then approve it, refine it, request product rework, or stop. Use as a Supervisor quality gate between product analysis and implementation and after implementation or testing results.
| name | product-audit |
| description | Inspect the current AxonFlow repository and runtime, identify evidence-backed product gaps, and turn the highest-value gap into one bounded implementation task. Use for product health reviews, self-optimization runs, backlog proposals, and deciding the next improvement before coding. |
Inspect before proposing. Read the recent optimization log, then use focused search to inspect only the product promise, implementation, tests, or runtime evidence needed for one candidate. Use references/audit-checklist.md to avoid shallow feature-only reviews. Stop inspecting as soon as one defect has two independent evidence points; use at most 12 tool calls.
Select exactly one task that is small enough for one Codex execution and can be verified automatically. Do not edit files. Do not propose broad rewrites, cosmetic-only work, credential changes, deployments, commits, or destructive operations.
Return only one JSON object:
{
"verdict": "proposed",
"task": "Concrete implementation instruction",
"problem": "Observed product problem",
"evidence": ["file:line, API observation, test gap, or UI observation"],
"requirements": ["Required behavior"],
"acceptance_criteria": ["Observable pass condition"],
"risk": "low|medium",
"affected_areas": ["repository-relative path or subsystem"],
"out_of_scope": ["Explicit exclusions"]
}
Use verdict=proposed only when the evidence is current and the acceptance criteria are testable. If no safe bounded task exists, return verdict=blocked and explain why.