Agent Skills
45 copyable SKILL.md files. Each follows the Agent Skills standard: a clear trigger, exact steps, and the pitfalls that bite. Drop any one into a Claude Code project.
45
Skills
8
Categories
SKILL.md
Standard
Skills of skills
Full guideHow to Write a Great Skill
The complete craft guide for SKILL.md files seeded from this repo's AGENTS.md: the four sections every skill needs, how to write a trigger an agent will actually match, what the pitfalls section is really for, and the signs a skill is complete vs still a draft.
MCP02Skill Testing Checklist
How to verify a skill works before shipping it: the five tests every skill must pass, how to run a trigger probe in Claude Code, how to spot a skill that is over-broad or stale, and the field test that proves it.
MCP03Skill Composition Patterns
How to build skills that work together in this repo: the entry-router, reference-file, and skill-chain patterns; split vs merge decisions; the versioning contract for repo skills vs flagship library skills; and safe cross-linking without duplicating AGENTS.md.
MCPWrite the trigger
The description is the whole match surface. Write it as a condition, not a topic.
Exact steps, real pitfalls
Show the actual command. State the actual failure. "Be careful" is not a pitfall.
Test before shipping
Probe the trigger, walk the procedure, audit each pitfall. Then watch a live agent use it.
Featured
45 in libraryAgent Fleet Coordination
Coordinate a fleet of subagents end to end: the orchestrator, finder, and verifier roles, single-writer file ownership, and a collect-then-apply harvest loop that merges results without collisions.
MCP Server Builder
Build a production MCP server as a Next.js route with mcp-handler: design tools as prompt engineering, wire strict zod schemas and Bearer auth, and evaluate every tool so the model actually calls it.
CLAUDE.md Authoring
Write a CLAUDE.md that actually changes agent behavior: the rules that are non-obvious, the commands that matter, and the pitfalls, without dumping the whole repo.
Credit Metering for SaaS
Meter AI features with a universal credit balance: per-action cost table, an atomic debit, owner bypass, and a clean 402 path that routes to pricing.
Verifying the Committed Tree Before Push
Catch the agent failure mode where a commit imports an uncommitted file: typecheck an isolated worktree of HEAD, not the dirty working tree, before pushing to main.
All skills
CommunityStructure App Router routes, server and client components, metadata, and data fetching the way Next.js 15+ actually wants.
Prove a UI change actually rendered by driving the app with Playwright: navigate, wait for real signals, screenshot, and assert on visible state.
Stand up a fast route-coverage suite in Playwright: assert every public page returns 200, gate the CI on it, and keep it deterministic with real waits.
Wire Clerk into a Next.js App Router app: provider, middleware, protected routes, and reading the user in server code without throwing.
Model a schema with Drizzle, generate and apply migrations against Neon Postgres, and keep local and production in sync without data loss.
Use the Neon serverless HTTP driver in a Next.js app that must build without a database: lazy connection, a guarded client, and pooled vs direct URL choices.
Survive domain-locked production Clerk keys in local dev: why auth() throws, the middleware bypass, and a getUserId helper that returns null instead of 500.
Issue and verify programmatic API keys without ever storing the secret: show once, persist only a sha256 hash plus a display prefix, verify by hashing the bearer token.
Diagnose the common ways a Next.js build passes locally but fails on Vercel: pnpm build gates, type errors, env vars, and uncommitted files.
Add oxlint as a fast, zero-error lint gate alongside a Next.js project: config, an autofix pass, and the errors-fail warnings-advisory split that keeps CI honest.
Gate agent-authored changes in CI: verify the committed tree not the working tree, reproduce the deploy's install and type gate exactly, and fail fast on drift.
Instrument a request with timed spans and one durable trace row: best-effort writes that never throw into the caller, per-thread grouping, and a waterfall the UI can render.
Catch the agent failure mode where a commit imports an uncommitted file: typecheck an isolated worktree of HEAD, not the dirty working tree, before pushing to main.
Coordinate a fleet of subagents end to end: the orchestrator, finder, and verifier roles, single-writer file ownership, and a collect-then-apply harvest loop that merges results without collisions.
Build a production MCP server as a Next.js route with mcp-handler: design tools as prompt engineering, wire strict zod schemas and Bearer auth, and evaluate every tool so the model actually calls it.
Expose a collection over MCP with the three-tier lean index -> manifest -> item pattern so an agent spends context only on what it opens. Covers each tier's payload shape, naming symmetry, and when to collapse the middle tier.
The complete craft guide for SKILL.md files seeded from this repo's AGENTS.md: the four sections every skill needs, how to write a trigger an agent will actually match, what the pitfalls section is really for, and the signs a skill is complete vs still a draft.
How to verify a skill works before shipping it: the five tests every skill must pass, how to run a trigger probe in Claude Code, how to spot a skill that is over-broad or stale, and the field test that proves it.
Prompt an LLM to emit a complete, self-contained HTML app and run it safely in a sandboxed iframe. Covers the output contract (one file, pinned CDN versions, module scripts), the sandbox and CSP for untrusted output, size caps, and the blank-frame failure modes.
Design LLM features that degrade gracefully when the model refuses, returns malformed output, or times out, instead of showing users a raw error.
Cut latency and token cost by structuring prompts so the stable prefix (system, tools, long context) is cached and only the tail changes per request.
Turn messy text into reliable typed data with an LLM: a strict schema, tool or JSON mode, validation, and a repair loop that never ships bad shapes.
Route every model through the Vercel AI Gateway with the AI SDK: one key, plain string model ids, a client-safe whitelist, and the gotchas that only bite in production.
Take a topic or a published video through research, drafting, SEO structure, and a pre-publish checklist to a shipped, source-linked, decision-intent post that ranks and stays fair to every tool covered.
Write comparison, pricing, and alternatives content that ranks and stays fair: real research, primary-source links, structured FAQ, and honest verdicts.
Build member and dashboard surfaces in the softer Vercel/shadcn idiom (rounded, neutral, generous spacing) while keeping them distinct from the hard-edged square marketing site. Covers tokens and component patterns.
Install shadcn/ui components and adapt them to a strict design contract: strip defaults that fight your system, keep the accessibility, ship consistent variants.
Render streaming LLM markdown safely in a chat UI with Streamdown: incomplete-token handling, a custom component map, and matching a strict design contract.
Entry skill for HeyGen's HyperFrames, a framework that renders video from HTML compositions. Routes any request to make, edit, animate, or render a video or motion graphic to the right HyperFrames workflow and domain skill.
Official best practices from the Remotion team for creating videos in React: project setup, animating with useCurrentFrame and interpolate, timing, assets, and rendering rules.
Monetize AI endpoints with a prepaid credit balance: the deduct-on-success ordering, an atomic single-UPDATE spend that cannot go negative under concurrency, a consistent 401/402/429 error envelope, rate limiting, an append-only ledger, and an owner bypass.
Meter AI features with a universal credit balance: per-action cost table, an atomic debit, owner bypass, and a clean 402 path that routes to pricing.
Wire Autumn (Stripe billing layer) for a universal credit balance: Clerk userId as the customer id, hosted checkout, Svix-signed webhooks, and grant-on-verify.
Let internal and founder accounts use a metered product for free without special-casing call sites: an isOwner flag read from the user row, honored inside the debit itself.

New tutorials, open-source projects, and deep dives on coding agents - delivered weekly.