plain-voice

Bucket: discipline · Invocation: automatic

Applies to every textual response the agent authors, including short conversational and status replies. It is not limited to formal prose artifacts.

Keeps written output carrying facts rather than the appearance of effort. Every sentence should leave the reader with a fact, an instruction, or a number.

Five recurring failures:

• Claims with no source or mechanism. Praise, promotional adjectives, attribution to unnamed experts.
• Length without meaning. Elevated synonyms, filler openers, stacked hedges, adverbs propping up a weak verb.
• Abstract metaphor in place of a concrete noun. Substrate, vector, surface, north star.
• Structure imposed on the content. Ideas forced into threes, "not just X but Y", bold labels that restate the line they introduce.
• Conversational residue. Congratulatory openers, closing offers of help, disclaimers about what could not be checked.

Two checks settle a doubtful sentence. Ask what it tells the reader to do or know; if it cannot be restated as a fact, an instruction, or a number, cut it. Then ask whether it could appear unchanged in another project's writing; if it could, it says nothing about this one.

Dashes do not belong in running text.

Language scope

The five failures apply to any language. Spelling, punctuation, and vocabulary tells are English and live in the skill's references/tells.md.

Established technical terms run the opposite way and stay in English whatever language surrounds them. edge case, pull request, and deploy are what a developer says out loud, while kantfall costs the reader a translation step before they can act on the sentence. Words with real native usage, such as databas or fil, translate normally. When unsure, keep the English term.

Boundary

Rewriting never reaches code, identifiers, structured data, quoted source material, or the developer's own prose. Leave those unchanged unless the request is to edit them.