agent-skills

Reviews and guides LLM/AI application engineering: prompt design, prompt caching, multimodal inputs, RAG, agent loops and tool design, resilience (rate limits, retries, fallbacks), memory, model migration, evals, testing, prompt-injection defence, and observability. Synthesises practices from Anthropic, OpenAI, Google, OWASP LLM Top 10, and practitioners (Hamel Husain, Eugene Yan, Chip Huyen). Triggers on "review my prompt", "design a system prompt", "optimise tokens", "set up RAG", "build an agent", "handle rate limits", "migrate to a new model", "write evals", "test my prompt", "audit AI code", "/ai-engineering".

Authors performant web animations CSS-first AND brainstorms the right feedback for an interaction via a verb-to-motion catalog. Covers GPU-safe properties, modern primitives (@starting-style, interpolate-size), View Transitions, scroll-driven timelines, state-choreography morphs, React state (Motion, AnimatePresence), advanced effects (Liquid Glass, glow, 3D tilt), external engines (Lottie, Rive), React Three Fiber, prefers-reduced-motion, AND perceived performance (skeleton loaders, optimistic UI, loader floor, predictive prefetch, stale-while-revalidate). Use when building transitions, hover effects, route changes, when an animation feels janky, when deciding what feedback an interaction should have, or when the app needs to feel snappier. Triggers on "animate this", "fade in", "hover effect", "liquid glass", "lottie", "rive", "ideal feedback", "how should X feel", "skeleton loader", "perceived performance", "feel faster", "optimistic UI", "/animations".

Authors performant React Native / Expo animations with Reanimated and gesture handling with react-native-gesture-handler, running motion on the UI thread so it stays at 60/120 fps. Covers worklets, shared values, useAnimatedStyle, withTiming/withSpring, layout animations, the Gesture API (pan/pinch/drag/swipe), Moti, Lottie, Rive, reduced motion, haptics, and UI-thread profiling. Use when building RN motion, wiring a gesture, or when an RN animation feels janky. For WEB animations (CSS, Motion, View Transitions) use the `animations` skill instead. Triggers on "reanimated", "gesture handler", "react native animation", "expo animation", "useSharedValue", "withSpring", "moti", "swipe gesture", "drag", "pinch to zoom", "/animations-native".

The phase-based machinery (0–7) behind the `aw` dispatcher — task intake through tested PR delivery in an isolated Git worktree, with optional companion skills for planning, quality gates, TDD, UX, code quality, docs, and CI verification. Companions never block; a skipped one is reported, never silent. NOT the entry point and not auto-triggered: a natural-language request to do work autonomously, end-to-end, in isolation, or in a worktree belongs to the `aw` skill, which detects the task tier and routes. Reach for this skill only to read or run the phase machinery directly, bypassing tier detection — invoke with /autonomous-workflow.

Ships autonomous, end-to-end coding work — implement a feature or fix, all the way to a tested draft PR — from a single opt-in entry point. Detects the task tier (Micro / Lite / Full) and routes: Micro/Lite run single-pass in this context; Full hands off to the aw-planner → aw-executor agents. Use when the user asks to do a task "autonomously", "independently", "in isolation", "in a worktree", "end-to-end", "all the way to a PR", to "ship this", "land this", "take care of this", or "handle this without me" — or invokes `/aw` directly. Opt-in, not a wrapper on casual edits; the routing rule's exclusion list governs when to hold back. Triggers on "implement autonomously", "end-to-end", "in a worktree", "ship this", "/aw".

Create the plan artifact (plan.md + checks.yaml executable acceptance checks) in `.agent/{branch}/` from the current conversation context. Captures all Phase 0-1 discussion into a structured, self-contained document that enables context recovery and session handoff. On every invocation, writes plan.md and re-derives checks.yaml from the Acceptance Criteria; an immutable plan.vN.md snapshot is written only in opt-in snapshot mode (`Skill("aw-create-plan", "snapshot")`). Use after planning is complete and confidence gate passes — and again on every plan iteration (user-requested refinement or Phase 4 auto-replan). Triggers on create plan, generate plan, write plan artifact, regenerate plan, iterate on plan.

Generate a walkthrough artifact (walkthrough.md) in `.agent/{branch}/` summarizing completed work for PR delivery. Gathers information from plan.md, git history, and test results to produce a comprehensive summary. Use at Phase 6 before creating the draft PR. Triggers on create walkthrough, generate walkthrough, write walkthrough artifact.

One-time (but safely re-runnable) setup flow that scaffolds a project's aw-tester aw-target: detects auth strategy, captures storage state, writes .claude/aw-targets/<target>.yml, and validates with a smoke spec. `--target local` (default) scaffolds the local dev target; `--target preview` scaffolds the PR-preview target ui-verify runs against (and is what `/ui-verify setup` delegates to). Also records the repo's UI surface (what counts as a UI change) so the is-ui-diff gate is accurate per repo, and the preview auth profile. Re-runs detect the existing aw-target and only re-prompt for what broke or changed. Triggers on "/aw-setup", "setup aw-tester", "scaffold aw-target".

Runs UI verification specs in-session through the claude-in-chrome extension — the Chrome sibling of the aw-tester agent. Reads the same specs.md (WHEN/THEN grammar or a Markdown intent spec) and aw-target.yml, drives Chrome interactively (navigate → read → act → assert, seeing the page between steps), and emits the same compact verdict block. Runs in the current session against an already-logged-in Chrome, so it is faster than the Playwright sub-agent for local runs, but it needs the browser extension and does not work in remote / CI envs. Falls back to aw-tester when the extension is not connected. Triggers on "run the spec in chrome", "verify with the chrome driver", "aw-tester-chrome", or a "ui-verify run --driver chrome" dispatch.

Applies behavioral science to make a desired behavior the easy default — for team adoption of tools, practices, and process changes as much as for product flows. Diagnoses why a behavior is not happening (COM-B, Fogg B=MAP), designs ethical interventions (EAST, choice architecture, defaults, friction, System 1 / System 2 fit), audits rollout plans, and builds behavior-change workshops and presentations. Never recommends a dark pattern or sludge. Use when the question is "why don't people do X" or "how do we get people to adopt X", not "is this component usable". Triggers on "behavioral design", "behavior change", "nudge", "drive adoption", "why aren't people using", "make it easy to adopt", "system 1 / system 2", "thinking fast and slow", "/behavioral-design".

Generates a personal markdown changelog of merged or closed pull requests authored by the current user and Linear tickets the user closed or worked on, over a configurable window (default 7 days), grouped by feature area (e.g. Dashboards, Agent0). Inputs sourced from `gh search prs --author=@me` and the Linear MCP. Use for weekly recaps, status updates, performance reviews, or end-of-sprint summaries. Triggers on "changelog", "what have I done", "weekly summary", "my recent work", "recap my week", "/changelog".

Selects the right chart type and visualization library for React/Next.js (web) and Expo/React Native (mobile) data visualization tasks, then applies accessibility and anti-pattern guardrails. Maps data shape and intent (comparison, composition, distribution, relationship, evolution, flow, geographic, hierarchical) to a recommended chart, then to a recommended library based on platform, dataset size, and design system. Defers cross-cutting UI concerns (contrast, touch targets, typography, copy) to the `ux` skill rather than restating them. Triggers on: "what chart should I use", "visualize this data", "build a chart", "build a dashboard", "data visualization", "pick a chart library", "graph this", "/charting".

Diagnoses and fixes failing CI, verifies locally, pushes, and confirms the relevant checks pass on the resulting revision. Uses an evidence-gated mechanical path for narrow failures and deeper diagnosis for ambiguous or shared changes. Never weakens checks. Currently supports GitHub Actions through gh or GitHub MCP. Triggers on "CI is failing", "fix the CI", "the build is red", "auto-fix this PR's checks", or "/ci-auto-fix".

Language-agnostic skill for authoring and reviewing code with low cognitive complexity, readability, and long-term maintainability. Covers guard clauses, single-responsibility functions, type-driven design (illegal states unrepresentable, branded primitives, discriminated unions), schema-first validation, single source of truth for union metadata, functional core + imperative shell, idempotency, and neighbour-pattern symmetry. Pairs with `tdd` for new code (rules apply in GREEN/REFACTOR). Stack-specific extensions live under `rules/stacks/<stack>/` (React, Next.js today; drop a subdirectory for any other language or framework). Use during PR review, after writing new code, in TDD GREEN/REFACTOR, or when asked to "improve quality", "make this readable", "reduce complexity", "deduplicate", "clean this up", or "/code-quality".

Rates confidence that the current work fully solves the stated requirement. Supports plan validation, code review, and analysis (root-cause, refactor, diagnose) modes. Plan mode combines LLM judgment with deterministic rule checks (multi-signal gate); a failed rule caps the gate at 89% regardless of LLM score. Use before committing to autonomous execution, after implementation, or during investigation. Triggers on "confidence check", "validate plan", "rate confidence", "quality gate", "/confidence".

Generates a short, narrative GitHub pull request description (≤ 25 lines, hard ceiling 40), pushes the branch, opens the PR as a draft, then runs review-loop (pr-reviewer → implement-suggestion → polish simplify, up to 5 iterations) until every review thread is resolved via fix or reply. Scale down with --no-review, --no-simplify, --quick (light mechanical pass only), or --no-quality (skip the loop). On a UI diff, injects a preview verification spec by default (--no-ui-verify to skip). Before the push, converges the branch by default with review-branch, which needs no PR, so the draft opens already review-clean (--no-pre-review to skip). With --split, breaks the branch diff into 2–4 focused, dependency-ordered draft PRs after user approval. Hands red CI to ci-auto-fix, which fixes or escalates. Invoke with /create-pr or /create-pr --split.

Scaffolds, reviews, upgrades, or diagnoses agent skills against best-practice frontmatter, progressive disclosure, token-aware structure, and the agent-skills.git symlink + inventory wiring. Modes: `scaffold` (default — new skill), `review` (audit existing skill), `upgrade` (split a single-file skill into multi-file), `diagnose` (retrospective failure analysis that emits a confidence-gated unified diff against any skill declaring a diagnostic surface). Triggers on "create a skill", "scaffold a skill", "new SKILL.md", "review this skill", "audit my skill", "upgrade this skill", "split this skill", "diagnose this skill", "why did the skill miss this", "/create-skill".

Adversarially challenges a proposed plan, code change, or bug diagnosis from a hostile pre-mortem perspective. Walks a fixed taxonomy of failure modes, blast radius, rollback, hidden coupling, and maintainability; every finding must cite a file, line, or named assumption; forces a steelman of at least one alternative. Surfaces concerns only — does not score (delegates to `/confidence`) and does not apply fixes. Use during planning before autonomous execution, before opening a high-stakes PR, or when a fix "feels off". One adversarial pass per run — naïve self-refine loops amplify bias. Modes: plan (default), code, analysis; add `deep` to run 3–5 independent persona lenses in parallel (sub-agents when available, personas found via `ideate`) and merge them — e.g. at the end of a feature. Triggers on "critical", "challenge this", "pre-mortem", "red-team this", "deep critical review", "review from every angle", "/critical".

Authors, audits, and maintains project documentation across CLAUDE.md / .claude/rules/, AGENTS.md, README.md, and Diátaxis docs/ trees (root + nested for monorepos). Four modes: init scaffolds a tiered docs setup from scratch; update detects drift (dead @imports, renamed commands, stale narrative) and incrementally refreshes via a Placement Resolver that pushes rules to the innermost-ancestor destination; readme writes or audits a README against the standard-readme spec; audit produces a documentation health report across every surface. Routes by kind: hard rules to CLAUDE.md, path-scoped patterns to .claude/rules/, narrative to docs/, marketing to README.md. Triggers on "init claude", "bootstrap docs", "scaffold CLAUDE.md", "update docs", "sync CLAUDE.md", "docs drift", "write a README", "audit our docs", "review the README", "Diátaxis", "/docs".

Developer Experience (DX) review and advisory skill for CLI tools, shell scripts, developer tooling, and automation. Analyzes code against established CLI design guidelines (clig.dev, Heroku CLI Style Guide, 12 Factor CLI), composability principles, error handling best practices, and developer ergonomics. Triggers on: "dx review", "review dx", "check cli", "improve the cli", "dx audit", "review this tool", "is this usable", "check ergonomics", "dx feedback", "review the script", "improve usability", "check error handling", "review output", "dx writing", "improve help text", "review flags", "make this more intuitive", "dx best practices", "/dx".