Style Guide as Prompt

# The style guide is the prompt

A common pattern for "make the LLM write in our voice" is to generate the text first and then run filters or rewriters on the output. The pipeline does the opposite. The full style guide lives inside the writer's system prompt. The model sees the rules before it generates a token, and most of the failure modes never appear in the output.

## What the style guide actually contains

The writing rules fall into four buckets.

**Forbidden words.** A list of corporate buzzwords and AI-tells that the model is told never to produce. Things like "leverage," "synergy," "cutting-edge," "innovative solutions," "robust," "seamless." If a sentence sounds like it could appear in a LinkedIn ad, it is wrong.

**Punctuation rules.** No em-dashes, no en-dashes, anywhere. Commas, periods, parentheses, or rewrites. Percents written with the symbol, not spelled out. Avoid colons and semicolons, default to periods and commas. These rules exist because the operator finds the LLM-typical punctuation distinctive and unwanted.

**Structural rules.** Length floor and ceiling at the letter level (~600 words). Sentence count per paragraph (5-6 sentences for paragraph one, 5-6 for paragraph two with longer word target, 5-6 for paragraph three). No three-line sentences anywhere. The greeting and the sign-off are fixed strings.

**Content rules.** Paragraph two only uses content from the selected fact bank, never the personal-anchor section. Paragraph three only uses one personal-anchor item, tied to a quote or named value from the company brief. Paragraph one introduces the candidate and the role through one storyline, not a list of facts. No claims about tools or experience that are not in the fact bank.

## Example rules in placeholder form