CatalystCode review for
your AI agent.
Catalyst is a Next.js boilerplate that runs seven checks after every Claude Code, Cursor, Codex, or OpenCode turn. Broken lint, types, dead code, duplication, and architecture get caught and fixed before the diff ever reaches you.
how it works
A review loop that runs itself.
- 01
Your agent finishes a turn
Claude Code, Cursor, Codex, or OpenCode edits your files and hands control back.
- 02
The stop hook fires
One runner kicks off all seven checks automatically. No prompt, no waiting on CI.
- 03
The diff gets gated
Pass, and it reaches your review clean. Fail, and the agent reads the error and fixes it, then runs again.
// the loop repeats until every check is green. you review working code, not typos.
the stop hook
Seven checks between your agent and your review queue.
Every one runs on every turn. Miss none, skip none, and the agent can't raise a threshold to make a failure disappear.
Lint and format, so style nits never land in your diff.
A full tsc --noEmit pass. A red squiggle never ships.
Flags unused exports, files, and dependencies the agent left behind.
Catches copy-paste. Fail it, and the agent must refactor, not raise the threshold.
Enforces this repo's own layer rules, not just generic lint.
Blocks circular imports before they turn into build-time mysteries.
Normalizes line endings, so diffs stay clean across machines.
Drop a script in .agents/hooks and add one line to the Stop array.
No magic. It's in the repo.
The whole pipeline is a handful of lines in a config file you own, mirrored for Cursor, Codex, and OpenCode. Read it, edit it, delete a check you don't want.
// .claude/settings.json
{
"hooks": {
"Stop": [
{ "hooks": [
{ "command": "bun .agents/hooks/check.ts" },
{ "command": "bun .agents/hooks/type-check.ts" },
{ "command": "bun .agents/hooks/knip.ts" },
{ "command": "bun .agents/hooks/jscpd.ts" },
{ "command": "bun .agents/hooks/klint.ts" },
{ "command": "bun .agents/hooks/madge.ts" },
{ "command": "bun .agents/hooks/lf.ts" }
] }
]
}
}.agents/skills
Playbooks your agent already knows.
The safety net stops bad edits. Skills push good ones, encoding the moves you reach for on every project.
Pulls in a high-quality UI design skill so generated interfaces don't default to purple-gradient AI slop.
Runs check-write, type-check, test, and build in order, auto-fixing what it safely can.
Scans for complexity, dead code, and performance issues, and proposes fixes one at a time, never a silent mass rewrite.
Reflects on what just went wrong and proposes a new rule for AGENTS.md, so the same mistake doesn't happen twice.
Teaches the agent to add or edit klint's own architecture rules, so the linter grows with your codebase.
Skills are plain markdown in .agents/skills, shared across Claude Code, Cursor, and the rest. Write your own playbook.
stock
Everything else is already wired.
The verification layer is the new part. The rest is the boring, essential plumbing you'd otherwise spend a week assembling.
Better Auth
Email and OAuth sessions, login, register, and password reset already built.
Prisma + PostgreSQL
A typed data layer with migrations and a clean DAO pattern.
tRPC + React Query
End-to-end typed API calls with a cache persisted to IndexedDB.
Stripe
Webhook handling wired in, ready for when you add checkout.
React Email + SES
Transactional emails authored in JSX and sent through Amazon SES.
Winston logging
Structured logs with levels, ready to point at a drain.
Biome + Husky
Formatting, linting, and pre-commit hooks configured from line one.
Docker + devcontainer
Postgres, Redis, and the app in one command, plus an agent container.