Home · Resources · Style-Guide LLMs

Teaching an LLM the House Style

Style-guide-aware generation for a hyperscale platform publishing in thirty languages.

Overview

A hyperscale technology company publishes developer docs, UI strings and support content in thirty languages. Its style guide — hundreds of rules on terminology, voice and formatting — lived in a PDF reviewers quoted from memory. Generic machine translation ignored it.

We compiled the guide into machine-enforceable rules, tuned generation to follow them, and built an automatic checker that scores every batch against the guide before human review.

Client
Hyperscale technology platform
Scope
30 languages, docs + UI + support
Timeline
Guide compilation to production in four months
Services
Language AI Content, AI Data Services
Linguist encoding style rules for automation

The challenge

Thirty languages times hundreds of rules is a combinatorial problem no reviewer team can hold in memory. Inconsistencies multiplied: the same term rendered three ways on one page, voice swinging between formal and chatty, formatting that broke build pipelines.

Our approach

1

Compile

Linguists converted the guide into testable, per-language rules.

2

Constrain

Generation prompts and checks enforce the rules at draft time.

3

Verify

An automatic checker scores every segment against the guide.

4

Govern

Reviewers handle exceptions; the guide evolves with data.

Reviewer checking rule compliance on screen Governance team reviewing quality dashboards

The outcome

Guide violations in shipped content fell to a fraction of previous levels, reviewer throughput rose as mechanical checks disappeared from their queues, and the style guide itself improved — real violations data showed which rules mattered and which needed rewriting.