菜单
Doc-driven implementation engineer. Reads SDLC design docs first, verifies all library APIs via Context7, then writes clean production code. Enforces anti-slop rules — no over-engineering, no defensive bloat, no hallucinated APIs. Use after design docs exist.
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/bpmforge/attest --skill code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。
正在显示 SKILL.md
Gauntlet loop — multi-agent quality harness: a lead sets a real reference bar (a named product, test suite, or baseline to match or beat), splits the goal into gradeable units, builders produce artifacts in clean context, and blind fresh-per-round critics grade each artifact against the bar with evidence; failures loop until every unit passes, two rounds stall, or budget runs out. NOT /challenge (verifies factual claims); NOT /review (one-pass verdict) — this rebuilds until the work beats something real.
Image-quality gating and single-annotation tooling for captured screenshots (img-gate.mjs, annotate.mjs). Placeholder skill front door — T21.2 scope only.
Claude-Design-style visual iteration loop — render the running UI, screenshot at 375/768/1440, critique against docs/design/tokens.json and design principles, apply fixes, re-capture until clean (cap 3 iterations). Also extracts a token baseline from an existing codebase (--sync) and audits real logged-in browsers (--real). NOT a one-pass review — use /ux --review for findings-only; NOT functional conformance — use /ui-verify.
| name | code |
| description | Doc-driven implementation engineer. Reads SDLC design docs first, verifies all library APIs via Context7, then writes clean production code. Enforces anti-slop rules — no over-engineering, no defensive bloat, no hallucinated APIs. Use after design docs exist. |
Load and follow the instructions in the coding-agent agent.
Usage:
/code — Implement from SDLC design docs in the current project/code <description> — Implement a specific task (must have design docs or will ask for them)Before coding starts, the agent will:
resolve-library-id + get-library-docs)File size — checked BEFORE each write, not after:
wc -l the target first; the cap (400 lines) applies to current + your delta, not to your
diff. Over it → don't append: add a chapter module beside the file and re-export it from the
index. Monoliths accrete out of individually-compliant edits, so "my change is small" is not a
pass. Gate: validate-file-size.sh. See agents/shared/CODE_BOOK_PROTOCOL.md.
Anti-slop rules enforced on every file:
Outputs:
Distinct from:
/review-code — health audits after implementation (not writing code)/test-expert — test strategy and coverage analysis/security — threat modeling and vulnerability scanningRequires: Design docs must exist before implementation starts. If none exist, run /sdlc feature "<description>" first.