Organisations with an established style guide try to apply it to generated text, and parts of it transfer cleanly while other parts do not. Which parts transfer follows a pattern that is worth knowing before writing the prompt.
Why a style guide suits a model at all
A style guide is an explicit specification of surface decisions, which is exactly the kind of instruction a model can act on.
Spelling conventions, date formats, capitalisation of product names, heading structure and citation format are all mechanical, and mechanical rules are followed reliably.
The guide also resolves the ambiguity that makes generated text feel generic, since a model without guidance defaults to an average of many conventions.
Applying a guide moves output away from that average towards something identifiably from one organisation.
This is one of the few interventions that improves both consistency and distinctiveness at once.
What survives and what does not
Rules with an unambiguous test survive. Use one spelling variant, write dates in a given order, never abbreviate a certain term.
Rules requiring judgement do not survive reliably. Write plainly, avoid jargon the reader will not know, keep the tone confident but not boastful.
The distinction is whether compliance can be checked without knowing the subject. If a checker needs domain knowledge, the model needs it too and usually lacks it.
Long guides also compete with themselves. A model given forty rules applies the early ones more consistently than the later ones.
Ordering the guide by importance rather than alphabetically has a measurable effect for this reason.
Why negative rules leak
Instructions phrased as prohibitions are followed less reliably than instructions phrased as requirements.
Naming a forbidden word raises its salience, and the word appears more often than it would have without the instruction in some cases.
Prohibitions also fail to specify a replacement, so the model must invent one, and the invented alternative may violate a different rule.
Rewriting a prohibition as a positive requirement fixes most of this. Instead of forbidding a construction, require the construction that should appear instead.
Where a prohibition is unavoidable, a post-generation check catches violations more dependably than the instruction prevents them.
How examples encode what rules cannot
Most style guides contain conventions their authors never wrote down, because they were absorbed rather than specified.
Typical unwritten conventions include how much context an opening paragraph assumes, whether examples come before or after the principle, and how the organisation handles uncertainty.
Approved documents carry all of this implicitly, and supplying two or three alongside the guide transfers more than adding rules does.
The examples should be genuinely representative rather than the best work ever produced, since the model will reproduce whatever is distinctive about them.
Pairing a compact rule list with real examples is more effective than either alone.
Where a checking pass beats a prompt
Some rules are better enforced after generation than requested before it.
Anything countable is a candidate: sentence length limits, banned terms, required sections, heading counts, and formats for numbers or dates.
A checking pass is deterministic, gives a clear list of violations, and does not consume space in the prompt that could carry more useful guidance.
The split that works in practice is to prompt for the rules that shape the writing and to check for the rules that merely constrain it.
This also produces a record of which rules are violated most, which is useful information about the guide itself.
Why terminology lists work best
Controlled vocabulary transfers better than any other part of a style guide.
A list of preferred terms with their forbidden alternatives is unambiguous, short, and easy to check, and terminology is where inconsistency is most visible to readers.
Product names, feature names, job titles and the organisation's own name in its various forms account for a large share of the errors that embarrass a team.
Including the list in every request costs little space and prevents the most conspicuous failures.
Where the list is long, supplying only the terms relevant to the current subject works better than supplying all of them.
How guides drift from practice
Applying a style guide to generated text exposes how far the guide has drifted from what the organisation actually publishes.
The model follows the written rule exactly, which surfaces conventions everyone abandoned years ago and rules that contradict each other.
This is uncomfortable and useful. A guide that produces obviously wrong output when followed literally was already producing inconsistent human writing.
Teams that go through this exercise typically shorten their guide substantially, discovering that many rules were never enforced.
The shortened guide then works better for both the model and the people, which is a rare alignment of interests.
What the guide cannot enforce
Style guides govern how something is expressed and say nothing about whether it is worth expressing.
Compliant output can still be inaccurate, unnecessary, badly argued or aimed at the wrong reader, and none of those failures registers against any rule.
There is a specific risk in this, because style compliance creates an impression of quality control that has not actually been applied to the substance.
Review processes that check the guide and stop there pass work that is wrong in ways the guide was never designed to catch.
Keeping the style check and the substance check as separate steps, done by different people or at different times, is the arrangement that prevents one from being mistaken for the other.