菜单
Define and maintain lightweight evaluations (acceptance criteria, regression checks, golden tests) so "green" is objective and repeatable.
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
Change-aware quality pass: diff scope, readability, function size and control flow, comments where non-obvious, and alignment with project linting. Use before merge or with handoff when the user wants industry-style maintainability beyond what tests assert.
Review code changes for correctness, security, project conventions, and maintainability. Use when reviewing a diff, before commit, or when the user asks for a code review.
Receiving-agent protocol. Boots a new agent into the minimum correct context (select + isolate) and produces a clear next-step plan without guessing.
Validates a game or interactive product for playability, clarity, feedback, and scale-readiness. Use at milestones, before large scope changes, or when the user asks if the product feels ready or is ready to scale.
Short-lived feature branches; TDD + lint + merge-ready command as exit criteria before commit; push and merge to main (or user-directed flow). Do not default to asking the user to open a PR. Use when implementing a feature or non-trivial fix, when the user asks for branch/git workflow, or after substantial edits that should not stay uncommitted.
Operating model for disciplined agent work: stay on scope, avoid guessing, decompose work into verifiable chunks, and keep context lean and durable.
| name | eval-engineer |
| description | Define and maintain lightweight evaluations (acceptance criteria, regression checks, golden tests) so "green" is objective and repeatable. |
Use this skill when:
This skill is tool-agnostic. It produces verifiable checks and a cheap evaluation loop.
Acceptance criteria (2–5 items)
Each AC must be verifiable. Avoid "works" or "seems fine".
Regression risk statement (1 paragraph)
What could break, and what signal will catch it?
Evaluation loop
Pick the smallest set that fits:
