Catch AI slop before you ship.
Design review on every commit, from the Taste Training and Agent Skills rules.
Install
npx taste-lint@latest init
Node 24.11+. Init adds a taste script that runs taste-lint lint --profile product.
Quickstart
npm run taste -- --dry-run
No key, no account. The mechanical checks run and fail the build on what they find.
To add the Jev review notes, get a Vercel AI Gateway key:
export AI_GATEWAY_API_KEY="your-vercel-ai-gateway-key"
npm run taste
Checks bill your key and cache between runs. --dry-run previews the cost first.
To block commits on the mechanical checks, add a pre-commit hook with lefthook, husky or lint-staged:
npx taste-lint init --hook lefthook
What it checks
- Motion:
transition-all and transition: all, ease-in, linear easing, transitions over 300ms, entrances from scale zero, framer-motion imports.
- Copy: the ghostwriter structure tells (participle tails, “it’s not X, it’s Y”, “Here’s the thing”), “click here” links, “successfully” toasts, emoji in UI, nominalisations, anthropomorphism, long sentences that chain clauses.
- Typography: straight quotes, three-dot ellipses, hyphens for dashes,
x for ×, missing space before a unit, light weights on small body text, uppercase without tracking.
- Code (
--profile code): N+1 queries, unbounded reads, blocking sync calls in handlers, tests that assert nothing or only that a mock was called, logic your coverage report shows no test runs, commented-out code, unused dependencies, dropped errors, secrets and personal data in logs, CI runs that never cancel.
- Search: pages with no metadata, a missing
metadataBase, sitemap or robots, noindex on public pages, canonical URLs on a dev host.
- Under review: vague errors, empty states with no action, bare “Confirm” labels, claims with nothing to check, missing error states. These go to Jev by TypeSafe AI, which answers with a probability.
28 checks block. 177 more report and never fail a run. They are pattern ports nobody has yet watched on a real codebase, and Jev questions no labelled corpus has yet promoted. See usage for how promotion works.
Profiles scope a run: product (tsx, jsx, css), writing (md, mdx, README), instructions (AGENTS.md, skills, plans), code (source, tests, CI), all.
Add your own voice with a rule pack: "rules": ["$GHOSTWRITER_HOME/taste-lint"] in taste-lint.config.json checks drafts against your ghostwriter profile.
Docs
License
MIT
Crafted by Matthew Blode
