A pattern library for keeping LLM-generated text aligned with a defined brand voice — across drafts, iterations, and team members.
LLMs default to generic institutional voice. Adjective-based instructions ("confident, warm, expert") don't fix it. A structured pre-flight block does — six fields placed at the top of every voice-sensitive prompt.
Most brand voice work is subtractive, not additive. The forbidden vocabulary list is the highest-leverage field.
This repo is for anyone who has tried to make a model write "in our voice" and watched it drift back to generic AI prose by the third sentence.
Large language models default to a kind of average institutional tone: warm, hedged, slightly American. That tone is fine — until your brand explicitly isn't that. If your voice is sharp, stripped, contrarian, or deeply local, every prompt is a small fight against the model's defaults.
Most teams handle this by:
- Pasting the entire brand-voice document into every prompt (works, but slow and expensive)
- Hoping a "use our brand voice" instruction is enough (it never is)
- Editing every output by hand (defeats the point)
There's a better pattern: structured pre-flight blocks.
- The Pattern — the core idea: a structured pre-flight block that goes before every voice-sensitive prompt
- Brand Voice Template — a fillable template for capturing your own brand voice in a model-readable format
- Examples — anonymized before/after prompts showing the pattern in action
- Anti-patterns — what doesn't work and why
- Solo operators writing in a defined voice (consultants, coaches, indie writers)
- Small teams trying to keep LLM-generated marketing text consistent
- Anyone who has read their own LLM output and thought "this could be from anyone"
If you have a 50-page brand guideline document already, this won't replace it — but it will make it actually usable in prompts.
Brand voice in prompts is not about adjectives. It's about constraints + counterexamples.
Telling a model "write in a confident, warm, expert tone" produces generic confident-warm-expert text.
Telling a model "do NOT use the words unleash, journey, holistic, transformative; do NOT begin with In today's...; one-line opening, no hedging" produces text that actually has a shape.
The pattern in this repo is built on this asymmetry: most brand voice work is subtractive, not additive.
CC BY 4.0 — use, adapt, remix freely with attribution.
- More before/after examples (different industries: SaaS, consulting, e-commerce, B2B)
- Domain-specific forbidden vocabulary lists (legal, medical, technical)
- Translation: German version (
pattern-de.md) - Companion tool: a small CLI that validates your generated text against your pre-flight block
Issues and pull requests welcome — see Contributing.
Real before/after examples from your own brand voice work are the most valuable contribution. See CONTRIBUTING.md for what we're looking for.
Built by Dirk Häger — independent learning architect at focusinstitute.io · LinkedIn
If this saves you editing time, ⭐ star the repo or share with your team.