Complete agent coding methods
These three collections cover different stretches of the work: deciding what to build, proving it works, and getting it through review.
- Lauren Tan’s Pstack — design and proof. Pstack gives an agent a way to design, implement, and show evidence for a change. Use
architectto sketch the types and module boundaries, andblast-radiusto trace what the change could break. For a flow with retries, races, or other hard-to-reach states, usecreate-formal-verification-skillto model the critical property andstatechartto make the flow visible. Usecreate-verification-skillto leave behind a repeatable way to exercise the real app. A model check and a run against the app answer different questions. - Matt Pocock — shared language and small slices. Matt Pocock’s skills start with agreement about the work.
grill-with-docspressures the plan and records domain terms and decisions as they settle.to-specandto-ticketsturn that agreement into a spec and small pieces of work.tddbuilds one slice at a time,diagnosing-bugsstarts by getting a reliable failure, andcode-reviewchecks the diff against both the spec and the repo’s standards. Reach forcodebase-designwhen the trouble is a module boundary. - Theo Browne’s tstack — review and delivery. tstack picks up when there is already a change or an inherited PR.
sol-loopalternates review and fixes;file-prframes the change for a reviewer;babysit-prfollows CI and review findings.modernizereassesses an old PR, whiletakeoverresumes someone else’s work. The bundled flows differ at the merge step.full-stopleaves the merge to you, andfull-sendincludes it.
Focused agent workflows
- Test audit. Check what each test protects, where to prove it, and whether another test already covers it.
- Eric Provencher’s Orchestrate. Split substantial Codex work into bounded scout and worker tasks while one coordinator owns decisions and integrates the results.
- Karpathy Guidelines. A short coding discipline derived from Andrej Karpathy’s observations: surface assumptions, keep the change small, and state what will prove it works.
- HumanLayer. Design an agent loop around the human review points, then package the repeatable part as a skill and monitor. I use
show-meall the time for a quick diagram or code sketch when prose is hard to scan. - Cursor Team Kit. Run a smoke test, turn a claim into a check with
verify-this, inspect a PR, or work through CI failures. The Cursor plugin catalogue also has separate plugins for iterative work, authoring plugins, and teaching. - Warp’s common skills. Connect product and technical specs to implementation and review. Its collection also covers merge risk, PR walkthroughs, and shareable readouts; some skills assume Warp’s own workflow.
- OpenClaw’s agent skills. Review a diff, run checks in a remote test environment, or turn an agent session into a shareable record.
- ast-grep skills. Find code by syntax rather than spelling. Use
ast-grep-outlinewhen you first need a structural map of a file. - ayghri’s I Have ADHD. Keep the next action visible and carry the state of the work across turns.
UI and design engineering
- Emil Kowalski. Design motion and gestures. Start with
animateorreview-animations; usechoreographwhen elements move together,gestureswhen motion follows the finger, andremove-animationswhen motion gets in the way. - Jakub Krehel. Improve the whole interface. Start with
better-interfacefor a ranked review, or pick the part you need: layout, type, color, copy, accessibility, or polish.build-designtranslates a design into existing components;state-machinelays out the visible states of a component so you can work on each one. - Benji Taylor’s Stageshow. Build a still or animated product mockup for a launch page, feature card, or walkthrough.
- Vojtěch Holík’s Good CSS. Find a native CSS solution before adding breakpoints, wrappers, or JavaScript.
- Florian Kiem. Use Prefer Container Queries for components that adapt to their own space, and WebGL Components for small shader visuals embedded in a real interface.
- shadcn/ui. Find and compose components from registries, or migrate a shadcn project from Radix to Base UI.
- Vercel’s agent skills. Implement view transitions or check a site against Web Interface Guidelines.
Runtime and framework skills
- React Doctor. Diagnose React problems, with separate skills for broader React cleanup, runtime performance, and Three.js.
- Kit Langton’s Effect skill. Design or review Effect v4 services, layers, schemas, streams, and tests.
- Vercel’s React skills. Review React and Next.js rendering performance, refactor component composition, or check React Native and Expo performance.
- Modern Web Guidance. Find native HTML, CSS, and browser APIs that fit your browser support targets. Install with
npx modern-web-guidance@latest install.
Hosting, deployment, and security
- Cloudflare’s skills. Choose a product, build and test a Worker, or work with Durable Objects, Agents, and Sandboxes. Use
cloudflare-onefor Zero Trust configuration andturnstile-spinfor bot verification. - Vercel’s deployment skills. Deploy an app or use project metrics to investigate performance and cost.
- Cursor’s Thermos. Use
thermo-nuclear-reviewfor a branch review focused on security and correctness.
Smart utilities and their companion skills
Use these when the agent needs to operate a browser, handle a capture, or collect evidence from a running app.
- Kit Langton’s utilities. Scrawl captures a narrated, marked-up screen recording for an agent to inspect. 2password lets a command use a 1Password secret without putting the value in the conversation. Motel collects local traces and logs; the companion
motel-debugskill tells the agent how to use them to investigate a running app. - Vercel’s agent-browser. A browser automation CLI for driving a flow and inspecting what actually rendered. Its companion skill teaches the agent the commands.
- Plannotator. A workspace for annotating plans and diffs. Its companion skills cover visual explainers, plan history, and setting up a goal.
- Guillermo Rauch’s UI recording timeline. Turn a screen recording into a scrubbable timeline of what appeared when, useful for studying a slow load or redirect.