Can an AI Model Obey a Strict Style Guide? I Tested Ten of Them | BymorningCan an AI Model Obey a Strict Style Guide? I Tested Ten of Them

Large language models write fluent text, but they do not obey rules well. Ask one for clear technical prose and you get text in the style the model prefers with a few small changes. That is fine for an automated LinkedIn slop post, but it is not fine for an aircraft maintenance manual, where a mechanic must understand only one thing from each sentence.

The aviation industry solved this long before AI with ASD-STE100 Simplified Technical English (STE), a rulebook of about 53 rules and 900 approved words. STE bans passive voice, "-ing" verb forms, and soft modal verbs, which are the fuel of natural English and prevalent in AI writing. Can an AI model obey a rulebook like that?

Anecdotally, there is some disagreement, so I built an experiment testing different approaches and models. The short answer is they can get some of the way with prompts, but full compliance requires checking against external tools.

Method

The test has four parts: a checker that counts rule violations, a fixed set of writing tasks, four prompt conditions that add effort step by step, and a feedback loop that sends the checker's findings back to the model. Each run is scored by violations per 100 words, as counted by the checker. Lower is better.

The checker

The checker ingests text and counts the number of violations of 10 rule families from ASD-STE100:

length of sentences

length of paragraphs

passive voice

"-ing" verb forms

perfect tenses

long strings of nouns

more than one instruction in a sentence

banned modal verbs

contractions

a list of unapproved words, with their approved replacements

The tasks

Each model got the same 10 tasks covering six domains, from aircraft maintenance to network hardware. Of the 10 tasks, 6 are procedures and 4 are descriptions. Six are rewrites of bad source text and four start from a blank page.

For example:

Rewrite the source text as a numbered, step-by-step removal procedure for the engine-driven hydraulic pump.

The prompts

Each task ran under four prompt conditions, so each model answered 40 prompts in total. The conditions add effort step by step:

C0: "write clear, accurate technical prose". The prompt does not name STE.

C1: one added sentence that names ASD-STE100. This measures what the model already knows.

C2: a one-page digest of the STE rules, in the prompt.

C3: the digest plus three before-and-after examples.

The feedback loop

The last part tests the "repair" step of the pipeline. The model gets its own text back plus the list of violations from the checker, with an instruction to remove the violations and keep all technical content. Each document gets up to three rounds, each in a fresh session, with no rules digest in the prompt. The loop starts from the C1 outputs.

Results: prompting alone

The table shows mean violations per 100 words, by condition. Lower is better. The sort is by C2, the best proxy for a basic Claude Skill or prompt.

Rank<br>Model<br>C0<br>C1<br>C2<br>C3<br>GovCloud

Claude Sonnet 5<br>8.00<br>2.17<br>1.66<br>1.53<br>yes

GPT-5.6 Sol (GPT flagship)<br>7.75<br>3.04<br>1.96<br>2.19<br>no

Claude Sonnet 4.5<br>9.61<br>3.34<br>2.34<br>1.94<br>yes

Claude Opus 4.8<br>7.43<br>2.51<br>2.36<br>1.69<br>yes

Claude Sonnet 4.6<br>7.13<br>5.36<br>2.38<br>1.95<br>no

Mistral Large 3 (675B)<br>9.40<br>4.29<br>2.58<br>2.02<br>no

GPT-5.6 Terra (GPT mini)<br>6.79<br>4.14<br>2.72<br>2.30<br>no

Claude Haiku 4.5<br>9.75<br>6.08<br>3.54<br>3.15<br>no

Llama 3.1 70B Instruct<br>11.50<br>5.79<br>4.98<br>4.09<br>no

10<br>Llama 3.1 8B Instruct<br>11.37<br>8.57<br>6.91<br>6.69<br>no

One sentence pays a large dividend. When the prompt only names the standard (C1), every model improves, with gains from 25 to 73 percent. So every model family knows something about STE from its education, but the depth varies by a factor of three. The one-page rules digest then cuts what remains by a third to a half for most models, and as prompt effort increases, the differences between the models shrink.

Claude Sonnet 5 is the best STE writer on the roster, though all frontier models perform very similarly. Three of the top four models are available on AWS GovCloud, and thus easily accessible to aerospace and defense teams with export controlled data.

More prompt is not always better, though. Nine models improved when I added the three before-and-after examples, but the GPT flagship model got worse: 1.96 to 2.19. My guess is that examples pull a model toward imitation of surface style, away from execution of rules.

The residual violations are not evenly spread across the rules. The "-ing" ban is very difficult to enforce: in 37 of the 40 model-and-condition cells, "-ing" verb forms are the top violation. However similar rules with a clearer trigger word collapse under a good prompt. Passive voice, for example, falls to zero by C3 for four of the five best models (Sonnet 5, the GPT flagship, Opus 4.8, and Sonnet 4.6 — though not Sonnet 4.5, which kept six), but "-ing" persists.

However, this metric alone does not...