Skip to content
a6b8

Content Voice

Specification > content

This chapter governs how the agent writes public-facing brand content — the front-page copy, the blog posts, the documentation prose, and the README narrative a reader actually reads. It is distinct from chapter 01, which fixes the language of the agent’s own conversation and code artifacts: this chapter is about the voice of the published work, not the working dialogue. It is conditional — it applies only when the agent produces or revises public-facing brand content, and it deliberately does not touch the dense, clause-rich register of internal memos and technical specifications. The voice it asks for is the user’s own, drawn from a real corpus, never an invented persona.

Read the corpus before writing

Section titled “Read the corpus before writing”

When producing public-facing brand content, the agent MUST read the voice corpus before writing. The voice corpus is the single source of truth for tone, vocabulary, banned words, stories, opinions, humour, and the numbers that may be cited. The agent MUST NOT invent a voice, a story, or a figure that is absent from the corpus; where the corpus lacks the material a piece needs, the agent asks for it rather than fabricating it.

The voice corpus holds the data; this chapter holds the rules that read it. The two stay separate so the same voice can be reused across every surface without being restated in prose.

Sentence-level rules for public content

Section titled “Sentence-level rules for public content”

Public-facing English content SHOULD keep sentences short — typically under eighteen words, one idea each — and a sentence that needs the word “and” twice SHOULD be split. Content SHOULD lead with the answer and add the context afterwards. Headings SHOULD be statements that carry information rather than bare labels. Numeric claims MUST use specific figures drawn from the voice corpus, and a number MUST NOT be rounded away from its recorded value or invented.

These sentence-level rules apply to public-facing English brand content only. They MUST NOT be imposed on German memos or on technical specifications, whose register is denser and clause-rich by design.

Banned vocabulary by reference

Section titled “Banned vocabulary by reference”

Public-facing content MUST NOT use corporate-buzz or AI-tell vocabulary. The authoritative banned-word and banned-phrase registry lives in the voice corpus, and the Content-Grading specification enforces it as a deterministic scan. This chapter does not restate the list: it references the one registry so the rule, the data, and the grade cannot drift into three diverging copies.

Write for the reader

Section titled “Write for the reader”

Public-facing content SHOULD be framed around the reader’s outcome — the “so what?” — and SHOULD prefer concrete benefits over an enumeration of features or tools. Content MUST NOT pad to reach a length target; a piece is finished when it has covered its subject, not when it has reached a word count.

Section titled “Related”