Use when the agent keeps making the same mistake across multiple PRs, when lint/test failures recur on the same rule, when PR review comments reveal a missing guide or unenforced convention, or when you want to proactively evolve the harness. Triggers on "self-evolve", "optimize harness", "update harness docs", "agent keeps making the same mistake", "harness self evolution", "harness 自我进化", "优化 harness", "更新 harness 文档".
Turn a single article, document, or web page into a standalone HTML learning artifact: either a bilingual digest or a chaptered Chinese study guide with source summaries. MUST use this skill — not raw summarization — whenever the user asks to summarize, digest, get key points from, extract takeaways from, or turn a URL, PDF, .docx, .md, .txt, paper, report, or blog post into notes or study material. Trigger on: summarize this, key points, takeaways, digest, highlights, TL;DR, what does it say, overview, 学习资料, 学习笔记, 复习资料, 总结, 摘要, 要点, 重点, 提取, 整理, 讲了什么. NOT for: writing new content, translating, comparing two documents, building apps, multi-source news, code review, or data/CSV analysis.
Use proactively when harness documentation needs updating — new module/package added, skill created or removed, CLI surface changed, exec plan completed, new convention established, or CLAUDE.md exceeds ~100 lines. Also use when asked to set up AGENTS.md/skills/sub-agents, fix agent drift, or retrofit guardrails.
Use when setting up a repository for autonomous coding agents, adding guardrails, context files, or automation so agents ship reliably without constant review. Triggers on "make this repo agent-friendly", "agents keep drifting", "set up AGENTS.md / skills / sub-agents", "harness engineering", architectural drift with agent-authored code, or retrofitting guardrails after output quality decayed.
Use when the user wants to resolve, fix, work through, or burn down open GitHub issues in a repository — "fix all the issues", "resolve these tickets", "work through the repo-scan issues", "clear the backlog", "fix issue
Use when the user asks to scan, audit, or sweep an entire GitHub repository for issues and file them as tickets — "scan this repo", "audit the codebase", "find bugs/security holes/missing tests", "check the docs are still accurate", "/repo-scan", "open issues for all the problems you find". Scans security, edge cases, testing coverage, and doc/code drift (CLAUDE.md, AGENTS.md, README.md and the docs they link to) without assuming business logic. Do NOT use for reviewing a single PR (use devpilot-pr-review) or language-specific style review (use devpilot-google-go-style).
Use when the user wants to review every pull request currently requesting their review, batch review a review queue, sweep pending PR reviews, or run PR reviews across all review-requested pull requests. Discovers the queue with `devpilot github prs review-queue` and delegates each real review to `devpilot-pr-review`; do not use for reviewing a single specified PR.
Use when the user wants to generate, scaffold, or auto-write end-to-end / browser tests for a web app — "generate e2e tests", "write Playwright tests", "auto-generate browser tests", "/e2e-tests", "test the login/checkout flow end to end". Playwright-first; drives the running app to capture real selectors and verifies each spec to green. Do NOT use for unit/integration tests, API-only tests, CLI tests, or non-Playwright e2e frameworks already present (it will stop).