taste

taste

Your code. Your taste.

taste makes every Claude edit match your house style — or Carmack's, or Linus's, or DHH's — so terse codebases stay terse and ornate ones stay ornate.

Install

claude plugin marketplace add Mapika/taste && claude plugin install taste@taste

No separate API key required — taste runs on your Claude Code subscription.

Pick a taste

Eight presets ship out of the box. One command switches.

Preset	Vibe	Best for
anthropic	Terse TypeScript, typed error hierarchy, full-word names, no barrel exports	TS backend work
carmack	Functions over classes, crash-on-failure, short names, zero ceremony	Performance-critical code, game engines, low-level libs
linus	Flat, boring, obvious. snake_case, no CamelCase, 8-space tabs, 80 cols, goto for cleanup	Systems code, kernel-adjacent work
dhh	Ornate, English-flavored DSLs, convention over config, metaprogramming is fine	Rails apps, writer-experience optimization
fowler	Long intention-revealing names, Extract Method, patterns by name, no anemic models	Enterprise OO, Java/.NET shops
dan-abramov	Hooks over classes, small composable functions, explicit data flow	React apps
37signals	Lean deps, narrow files, human copy, good READMEs	Small-team OSS, writer-friendly code
google-go	Short names in short scopes, errors as values, gofmt, table-driven tests	Go services

/taste:use carmack      # switch active profile
/taste:learn            # distill a profile from THIS repo

How it works

taste wires up three Claude Code hooks:

1. PreToolUse (on Edit/Write) runs your profile's regex and file-name rules inline. <1 ms, never blocks. Flags banned tokens and — when a rule declares an explicit fix: replacement — rewrites before the edit lands.
2. PostToolUse spawns a detached worker that calls claude -p with your profile as the system prompt, gets a verdict from Haiku (escalating to Sonnet on low confidence in strict mode), and appends to .taste/audit.jsonl. Takes ~10–15 s — but detached, so the user is never waiting.
3. UserPromptSubmit reads fresh audit entries and emits them as additionalContext on your next turn. Claude sees ## Taste verdicts (background, 2 new) at the top of its context and can act on them.

A SHA-256 verdict cache in .taste/cache.json means identical edits skip the LLM entirely. /taste:stats summarizes the session. /taste:diff lists the last 20 rewrites. /taste:misfire flags a false fix for the profile gallery.

Modes

/taste:set-mode async       # default — T1 inline, T3 in background
/taste:set-mode strict      # T1 + T3 both inline; user waits ~15 s per edit
/taste:set-mode score-only  # regex rules only, no LLM calls ever

• async is the default. Edits land instantly; LLM verdicts surface on the next prompt.
• strict blocks on the LLM judge so rewrites apply before the edit commits. Slow, but catches fuzzy violations pre-landing.
• score-only is a hard-rule firewall with no LLM cost.

Cost

On a Claude Code subscription: $0 — all calls route through your plan.

Direct-API billing: a Haiku judge runs ~$0.005–$0.01 per edit. Typical 50-edit session is ~$0.25–$0.50. The verdict cache means repeats are free. Track per-session cost with /taste:stats.

/taste:set-budget <usd> records a budget intent in .taste/config.json. Enforcement (auto-downgrade to score-only when hit) is not yet wired — use /taste:stats to watch spend for now.

Does it work?

Evaluated against 400 hand-curated fixtures (50 per preset) — should-fix, should-pass, edge, and adversarial. Judge runs Haiku 3-vote majority; verdicts below are verdict-accuracy.

preset	overall	should-fix	should-pass	false-fix
anthropic	98%	100%	100%	0
dhh	96%	100%	100%	0
google-go	94%	85%	100%	0
fowler	88%	75%	93%	1
linus	84%	80%	93%	1
carmack	80%	80%	100%	0
dan-abramov	78%	50%	100%	0
37signals	76%	65%	100%	0
Aggregate	87%	79%	98.3%	2

Trust-critical number: 0.5% false-fix rate (2 of 400). If your code is already clean, taste almost never rewrites it wrong.

Run the eval yourself: node dist/eval/harness.js <preset>. Reports drop into eval/reports/.

Add your own preset

See presets/SOURCES.md for the template. The quality bar:

• Cite ≥20 real snippets in presets/<name>.corpus.md with URLs
• Hit ≥95% style-compliance on your fixture set
• <5% false-fix rate on should-pass fixtures

License