GITHUB AGENT SKILLS / 技能收录
给你的 AI 助手
装上专业技能
我们从 GitHub 持续收录、审核并分类优质的 Agent Skills，按职业与使用场景整理，复制一行命令即可安装到 Claude Code、Cursor 等智能体。
按职业浏览 Agent Skills
为不同岗位挑选的高频技能
按使用场景浏览 Agent Skills
从你要完成的任务出发
按创作者浏览 Agent Skills
按 GitHub Stars 排序的优秀创作者
Microsoft
@microsoft
openclaw
@openclaw
OpenAI
@openai
The Apache Software Foundation
@apache
GitHub
@github
Anthropic
@anthropics
AutoGPT
@Significant-Gravitas
Garry Tan
@garrytan全部 Skills 192262
explain
Explain WHY a piece of code is structured the way it is. Synthesises architectural-memory similar-symbols, git history (blame + commits + PR context), AGENTS.md/CLAUDE.md principles, and any plan documents that mention the target. Useful for onboarding ("why is this here?"), debugging ("why is X structured this way and not Y?"), and refactoring ("can I change this safely or is there hidden context?"). Pick by INPUT: a TOPIC to survey comes here ("have we touched rate limiting?"); a CLAIM to test goes to /investigate ("that landed in July" — it uses --history as one of its instruments). Invoke on the question's shape, don't pre-check the target exists first — Step 0 validates it and reports cleanly if missing; don't Grep/Glob and answer directly instead. Triggers on: "why is this", "explain this code", "why does this exist", "what is this for", "/explain", "give me context on". Full command syntax: see the Usage section in this skill.
cycle
End-to-end feature cycle orchestrator. Runs the full skill chain in sequence: /plan → /audit-plan → (wait for human implementation) → /audit-code → /persona-test (if frontend/full-stack) → /ux-lock (if fixes shipped) → /ship. Use when starting a new feature or non-trivial fix and you want the whole workflow on autopilot. Supports resuming from an existing plan or straight to code-audit, per-step skips, max-round pass-through, an opt-in autonomous mode that implements + audits each plan cluster (the default still pauses for the human), and a remediate mode that fixes triaged persona-test/audit findings against an accepted plan with a focused audit and a rerun of only the affected journeys. Triggers on: "run the full cycle", "do the whole flow", "plan + audit + ship", "feature cycle", "remediate the persona findings", "/cycle". Full command syntax: see the Usage section in this skill.
click-test
Structural DOM audit of a live app — walk every interactive element and assert semantic-HTML contracts (duplicate IDs, orphan labels, inputs without names, ARIA misuse, heading hierarchy, missing alt text, undersized touch targets). Complements /persona-test: persona-test catches narrative UX issues a real user would hit; click-test catches structural issues that hide in JS-rendered surfaces and silently break assistive tech, form submissions, or React reconciliation. Drives a browser via the shared driver contract; optional --with-modals opens each modal/dropdown and rescans; device presets and matrix mode cover responsive breakpoints. Triggers on: "click test", "click-test", "structural audit", "DOM audit", "accessibility audit", "duplicate ID check", "walk every element", "audit the DOM", "structural QA", "/click-test". Full command syntax: see the Usage section in this skill.
brainstorm
Multi-LLM concept-level brainstorming. Sends the user's topic to OpenAI (and optionally Gemini) so the user can compare independent perspectives alongside Claude's. Convergence is manual — Claude waits for the user to ask for synthesis instead of auto-merging the views. Supports model selection, repo-architecture context attachment (auto on arch topics), a debate second round, depth budgets, session resume, extra context and artifact attachment, and saving keeper insights from prior rounds. Triggers on: "brainstorm", "let's think about", "get other LLMs on this", "what would Gemini/GPT say", "/brainstorm". Full command syntax: see the Usage section in this skill.
audit-plan
Iteratively audit a plan file (docs/plans/*.md) with GPT + Gemini final gate. Multi-round refinement with an acceptance-rate early-stop (3 rounds default). Ledger-driven suppression on R2+ rounds prevents finding churn. Judges the plan DOCUMENT before anyone builds it. "Was the plan actually built?" is /ux-lock verify — that reads a live app, this reads only the file. Triggers on: "audit the plan", "iterate on the plan", "plan-audit", "audit docs/plans/", "review the plan", "is this plan good", "check the plan", "/audit-plan". Usage: /audit-plan <plan-file> — Audit existing plan iteratively Usage: /audit-plan <task-description> — Generate plan with /plan, then audit
audit-code
Iteratively audit code against a plan with GPT + Gemini final gate. 5-pass parallel static analysis (structure, wiring, backend, frontend, sustainability), R2+ ledger-driven suppression, debt capture for out-of-scope findings, max 6 rounds with 2-stable-rounds convergence. Invoke this for ANY code-correctness review, even a small pasted snippet or diff — don't just review it yourself directly; a single read-through is not a substitute for the multi-pass GPT+Gemini adjudication this skill performs, and skipping it means the review didn't actually happen. Triggers on: "audit my code", "audit my changes", "audit my PR", "check the implementation", "review my code", "audit code", "/audit-code", "audit this", "verify the implementation". Usage: /audit-code <plan-file> — Audit code against plan (--scope diff default) Usage: /audit-code <plan-file> --scope plan — Audit only files mentioned in the plan Usage: /audit-code <plan-file> --scope full — Full repo audit
ai-context-management
Manage AGENTS.md and CLAUDE.md alignment across a repo so Claude, Copilot, Cursor, Windsurf, and other AI agents share the same project context. Detects drift, reconciles manually-introduced divergence, and migrates legacy CLAUDE.md-canonical repos to AGENTS.md-canonical. Triggers on: "audit AGENTS.md", "check context drift", "AGENTS.md vs CLAUDE.md", "fix context drift", "migrate to AGENTS.md", "context sync", "what tools read which file", "ai-context-management". Usage: /ai-context-management audit — Score drift + report findings Usage: /ai-context-management reconcile — Apply patches to fix drift Usage: /ai-context-management migrate — Flip legacy repo to AGENTS.md canonical
security-strategy
On-demand maintenance of the per-repo security memory: bootstrap an initial threat model + add/append incidents to docs/security-strategy.md with proper marker comments, then log the operation to the Postgres security_incident_log audit table. NEVER gates planning — /plan reads docs/security-strategy.md directly via Phase 0.5b. Triggers on: "security strategy", "add security incident", "draft incident", "/security-strategy", "bootstrap security memory". Usage: /security-strategy bootstrap — interview to seed initial threat model + first incident /security-strategy add-incident — interactive draft of one new incident entry /security-strategy add-incident from-commit <sha> — pre-fill from commit subject + changed files
visual-audit
Math-first, deterministic visual/paint inspection — the 4th UX lens, complementing persona-test (journey), click-test (page), and nav-audit (system). Drives Playwright + computed styles + bounding boxes + CDP forced-pseudo-state to audit declared-token conformance, light/dark theme parity, layout physics, and affordance signifiers. A VLM is advisory-only and never gates. CI gate is drift-only on the changed contracted surface. Triggers on: "visual audit", "check styling consistency", "theme parity", "dark mode parity", "design token drift", "paint audit", "button/card consistency", "focus ring audit", "/visual-audit".
ux-lock
Generate Playwright e2e specs. Two modes: 1. LOCK mode — pin a fix's DOM contract so it doesn't regress (default). 2. VERIFY mode — check that a /plan plan was actually implemented by parsing its Acceptance Criteria section and driving the live URL. VERIFY mode grades a plan against a RUNNING app. To judge the plan document itself, before anything is built, that is /audit-plan. Triggers on: "ux lock", "lock ux", "lock in the fix", "write regression spec", "generate e2e test for", "regression test for commit", "lock this fix", "verify the plan", "verify plan implementation", "check the plan was built", "audit frontend implementation", "did we ship the plan". Full command syntax: see the Usage section in this skill.
skills
Quick reference for every available skill in this repo — name, one-line purpose, triggers, usage examples — without having to open each individual SKILL.md document. Triggers on: "/skills", "list skills", "what skills do you have", "show all skills", "skill cheatsheet", "skills help", "show me the skills", "what can you do". Usage: /skills — Compact reference table for all skills /skills <name> — Full detail for one skill (e.g. /skills explain) /skills --search "<term>" — Filter by name/description match
ship
Sync all project documentation, optionally update a plan, then commit and push to git. Updates status.md (session log), syncs CLAUDE.md to AGENTS.md, and handles git workflow. Use when the user is ready to commit and push their work. Recommend it on: "ship it", "commit and push", "push my changes", "ready to ship". DO NOT INVOKE THIS SKILL ON YOUR OWN INITIATIVE — it commits and pushes, which is not undoable once the push lands. Run it only when the user asked for it in their own words this turn. If they merely approved a design ("looks good, ship it" about a plan), say /ship is available and stop. Their invocation IS the approval that lets every step skip confirmation, which is exactly why it cannot be self-invoked. Full command syntax: see the Usage section in this skill.
security-strategy
On-demand maintenance of the per-repo security memory: bootstrap an initial threat model + add/append incidents to docs/security-strategy.md with proper marker comments, then refresh the Supabase index. NEVER gates planning — /plan reads the resulting memory automatically via Phase 0.5b. Triggers on: "security strategy", "add security incident", "draft incident", "/security-strategy", "bootstrap security memory". Usage: /security-strategy bootstrap — interview to seed initial threat model + first incident /security-strategy add-incident — interactive draft of one new incident entry /security-strategy add-incident from-commit <sha> — pre-fill from commit subject + changed files
plan
Unified architecture + UX planner with engineering principles. Auto-detects whether the work is backend-only, frontend-only, or full-stack and applies the appropriate principle sets. Use when the user asks to plan, design, or architect anything — features, refactors, API endpoints, services, UI components, pages, layouts, user flows, modals, forms, visual changes, database schemas. Also auto-invoke when detecting planning context like "I want to add", "let's design", "plan the implementation", "how should we structure this", "I need to refactor", "design the UI for", "build a component", "improve the UX of". Invoke on the request's shape; don't explore the repo for stack context first and answer directly — the skill gathers that itself once invoked. Accepts a task description as argument; an explicit scope hint and diagram controls are available as flags. Triggers on: "plan", "design this", "architect this". Full command syntax: see the Usage section in this skill.
persona-test
Persona-driven exploratory browser testing against a live URL. Drives a browser as a specific user persona using a Plan → Act → Reflect loop, takes screenshots at each step, analyses UX and functional issues with confidence scoring, and returns a structured P0–P3 severity report plus a qualitative persona debrief. Personas are tracked per app URL; pair mode runs two opposed personas back-to-back; consistency mode runs a deterministic code-driven canary. Use for exploratory QA against deployed apps — not scripted regression tests. Invoke on the request's shape; don't check for browser tools first and decline — the skill resolves its own driver (Playwright MCP, host tools, BrightData, or read-only) once invoked. Triggers on: "persona test", "test as", "explore the app as", "run persona test", "test the site as", "browse the app as", "QA as", "list personas", "add persona", "who are my personas", "which persona should test". Full command syntax: see the Usage section in this skill.
nav-audit
Static, code-derived navigation / information-architecture audit — the system-level third lens complementing /persona-test (journey-level) and /click-test (page-level). Builds the whole nav graph from source, checks whether what's OFFERED matches what's NEEDED (grounded in the persona registry), and gates CI only on declared-intent regressions. A verify mode drives the live app to reconcile static-vs-live (per-persona scorecard + live findings); a bootstrap mode emits a review-queue nav-contract.json skeleton on first run. Triggers on: "map the nav flow", "navigation audit", "IA audit", "is the menu coherent", "nav map", "information architecture", "offered vs needed", "/nav-audit", "/ia-map". Full command syntax: see the Usage section in this skill.
investigate
Settle a factual claim about this repo with a measurement, not a recollection. Use when a number, a date, or a current state is asserted and needs checking: verify a figure in a doc or plan, establish when a change actually landed, reproduce a past measurement, or confirm whether something is really enabled/disabled/enforced. Forces a cross-check by a second INDEPENDENT method and a positive control, because a zero result and a broken instrument look identical. Pick by INPUT: a CLAIM to test comes here ("the cache cut cost 40%", "that landed in July"); a TOPIC to survey goes to /explain --history ("have we touched rate limiting?"). "Why is it this way" is /explain. Triggers on: "verify that", "is that number right", "is this still true", "when did we actually", "did that change land", "reproduce the measurement", "check whether X is really", "confirm the claim", "/investigate". Full command syntax: see the Usage section in this skill.
fleet
Coordinate SEVERAL CONCURRENT AI coding sessions (desktop chips, linked worktrees, a shared tree): see who is doing what, stop two sessions building the same thing, and land ready branches with ONE combined test run after explicit human approval. /fleet COORDINATES CONCURRENT SESSIONS; it does not plan (/plan), audit (/audit-code) or ship a single change (/ship), and it is not /cycle (one feature, one session) nor /brainstorm. Coordination is cooperative, not enforced. Triggers on: "fleet", "what are my sessions doing", "status of all branches", "spin up sessions", "land these branches together", "/fleet". Full command syntax: see the Usage section in this skill.