taste-lint

Is: the operator's guide to taste-lint, an npm CLI that turns copywriting, typography, interaction and motion rules from mblode/agent-skills and mblode/taste-training into checks over Markdown, MDX, TSX and rendered pages. Mechanical where a regex or a real value decides; a calibrated probability from TypeSafe Jev where a judgement is needed.

Is not: the rules themselves (they ship inside the package and trace back to the skill or lesson they came from), a copy rewriter (docs-writing, ui-design), or a replacement for the typography-audit and ui-design audits.

Commands

taste-lint schema prints every command, flag and default as JSON; --help is the human form. The ones the workflow turns on:

Command	Use
taste-lint lint --profile product --dry-run	Every mechanical check, no key, no calls. Act findings already fail here. Always first.
taste-lint lint --profile product	The run with Jev review notes. Needs the user’s AI_GATEWAY_API_KEY (Vercel AI Gateway). Results land in results/lint-<timestamp>.json; answers are cached under results/cache, so a rerun over unchanged files costs nothing.
taste-lint lint --profile code	Source, tests and CI: performance, data access, test value and gaps, dead code, telemetry, sensitive data. Every rule here reports and never fails a run.
taste-lint lint --since origin/main --output sarif	Only findings on changed lines, for code review tools.
taste-lint lint <paths> --output json	For scripts: findings, scorecard and usage on stdout, and an error as a { error, code, message } envelope instead of text. --output sarif for code scanning.
taste-lint lint --url <url>	Computed styles through style-capture and a local Chromium; --selector narrows the page.
taste-lint eval and taste-lint tune	Precision, recall and calibration per rule on the labelled corpus; tune --write is the only thing that promotes a rule.

Node 24. Jev input costs $0.042 per million tokens and output is free; the summary line prints the actual cost after every run.

Reading a finding

Every finding carries two independent facts. Severity (major, minor) says how bad it is if real and comes from the rule. Band says how sure the tool is: act (a mechanical hit, or Jev at or above the rule's act threshold; fails the run), review (between the review and act thresholds, or any hit from a review-only rule; printed with a ?, never fails the run), and silent. A [MAJOR?] is a major if it is real and the run did not fail on it; say both halves and never lower the severity to express the doubt, which the band already carries.

unknown is a third outcome, not a pass: a rule could not decide because a value was unresolved (a Tailwind theme token, a missing class list) or a precondition was unmet. --verbose lists them.

Workflow

Done means: the touched files re-linted with no act-band findings left that you have not either fixed or explained, and a summary that quotes units, act, review, unknown and cost from the last run plus the scorecard delta by category.

Linting reads files and writes only under results/, and a whole repository costs cents, so run --dry-run and the lint without asking. --fix rewrites the prose ranges of act-band findings in the working tree (curly quotes, ellipses, multiplication signs, unit spaces) and nothing else; run it, then show the diff.

1. --dry-run; quote the request count and cost.
2. Run the lint. Fix act findings in the copy or the classes, not by editing the rule.
3. Re-run only the files you touched; cached answers make the rest free.
4. A wrong finding is corpus evidence: add the unit to the corpus with the correct label rather than lowering a threshold by hand; tune decides thresholds.

Gotchas

• Projects that curl quotes at build time (remark-smartypants in next.config.*) are detected; the straight-quote rule then skips Markdown and still checks JSX. Force it either way with "smartQuotesAtBuild" in taste-lint.config.json.
• Deliberately bad copy (course stimuli, Incorrect examples, sample text) is corpus, not a lint target. Exclude it with --exclude or the config file, or suppress one unit with taste-lint-ignore: <rule-id> on the line above; ui-audit-ignore:<id> from ui-design is honoured for ported rules.
• Most of the pack is review-only (ported patterns until someone has watched them on a real codebase, Jev rules until tune --write promotes them). The report counts those notes by rule and lists them only with --verbose. Read the act band first; treat the notes as a reading list, not a backlog.
• Class-list rules see only static Tailwind classes. Theme tokens, template expressions and computed values are unknown until you run --url.
• Never paste a raw provider response or error body into an issue or a summary; the tool records categories and token counts only, and so should you.

Docs

https://blode.co/taste-lint/docs

Related skills

From mblode/agent-skills: docs-writing and ui-design own the wording of a fix; typography-audit is the type review this tool's typography pack is harvested from; ui-design and ui-animation supply the interaction, craft and motion checks.

Maintenance only: evals/evals.json holds the behavioural scenarios and routing prompts for anyone changing this skill. It never loads during a user task.