Voices: AI output styles powered by Vale

Writing voices from the output-style catalog, written as Vale rules instead of prompts. Checked on every draft rather than remembered, and free until something breaks one.

Demo	Briefs	Agent tools	Install Vale

📦 Install

Voices requires Vale v3.20.0 or later, and builds on Std, the standard library its length rules extend. Sync pulls Std in automatically; it stays present without being enabled.

Packages = https://github.com/jdkato/voices/releases/latest/download/Voices.zip

[*.md]
BasedOnStyles = Voices, Direct

$ vale sync

Voices is the shared core and is always on. Add one voice beside it.

Voice	From	Adds
Voices	no-ai-slop	Inflated words, binary contrasts, throat-clearing, puffery, weasel attribution, colon reveals, recap endings, weak verbs
Direct	no-slop	No hedging, no preamble, sentences under 25 words
GenZ	gen-z, street	One slang term a sentence, two a paragraph, at least one
Coach	coach	One note (100 words), one image, one "Next:" action
Simple	thing-explainer	Only the 850 words of Basic English

The catalog holds dozens of styles; these are the ones that convert into rules a linter can hold. Each covers a different kind of constraint -- patterns, register, structure, vocabulary -- so they compose instead of overlapping.

Coach wants Voices.ColonReveal = NO in your config: its required Next: label is the colon reveal the shared core forbids.

See vale.sh/blog/voices for one draft run through each voice.

🎛️ Tune a voice

A voice is a named setting of dials, and the dials are addressable. Levels and toggles from your config, as always; scalars with the bracket key:

[*.md]
BasedOnStyles = Voices, Direct
Direct.Length[max] = 30
Voices.ColonReveal = NO

To change what a rule matches, extend it in a style of your own -- your name on the alert, the machinery inherited:

# styles/House/Hedging.yml
extends: Direct.Hedging
message: "We don't hedge."
level: error
tokens+:
  - perhaps the best answer is

[*.md]
BasedOnStyles = Voices, Direct
House.Hedging = YES
Direct.Hedging = NO

The pair of lines says exactly what happens: your rule in, its parent out. The disable matters -- the child matches everything its parent does, and with both enabled every hedge would be reported twice. Leave both on only when that's the intent: a broad parent at suggestion under a narrowed child at error is a legitimate shape.

🤖 Use it with an agent

Install the Vale agent tools plugin:

/plugin marketplace add vale-cli/agent-tools
/plugin install vale@agent-tools

Every prose file Claude writes is linted, and the alerts go back into the same turn. Rules with one right answer carry the replacement, so those are applied without a decision.

Tip

The hook relays errors only. Simple is advisory by design. Widen the hook's level in /plugin, or raise a rule with Simple.Vocabulary = error.

To prime the model as well as check it, briefs/ holds a generated summary per voice: cat briefs/Core.md briefs/Direct.md >> CLAUDE.md.

Outside Claude Code, Vale reads stdin and sets an exit code:

$ vale --ext=.md --output=JSON < draft.md

❤️ Credit

The core's word lists and patterns come from no-ai-slop by Peter Yang, MIT. The voices follow awesome-claude-output-styles by Serge Shima, also MIT. Simple uses Basic English, C. K. Ogden's 850-word vocabulary from 1930.

NOTICE has the full attribution, and ships inside the archive.

🧪 Tests

$ ./test.sh

One draft against every voice, compared to a golden file; then each rewrite, required to produce nothing. A Vale rule that matches nothing fails silently, so the paired fixture is what tells "no violations" from "no rule". test.sh expects a Std checkout beside this repository, or STD pointing at one.

License

MIT.