Sansi, have you noticed how much filler AI replies carry these days?

Is there a cure for that?

There is, and the prescription is over 40 years old.

It's called aircraft maintenance manual style.

That name sounds like it has nothing to do with AI.

It was born before modern AI existed.

In the 1970s, aircraft maintenance crews in Europe came from many countries. Most of them were not native English speakers, and one misread sentence could get a component installed wrong.

To help maintenance workers read the manuals, the industry drafted a set of English writing rules: which words you may use, how long a sentence may run, all of it fixed.

That controlled English later became the ASD-STE100 standard.

So how did that standard end up anywhere near AI?

Around mid-year, someone turned ASD-STE100 into an open-source Skill.

In early October, Karpathy shared the trick on X: just tell the model to write replies in aircraft maintenance manual style.

What's the secret of aircraft maintenance manual style? How does it stop AI from rambling?

Four rules do most of the work:

  1. One action per sentence.

  1. Keep a procedural sentence to 20 words or fewer.

  1. Use the active voice and name the actor.

  1. Use one name for one concept, from start to finish.

Not even synonyms? I used to think swapping words was good writing.

That's its limit: it doesn't aim for beautiful prose, so it's no good for essays.

It fits the times when you don't want to think and just want to follow the rules mechanically.

Picture a technician holding a flashlight in one hand and a wrench in the other. He has no attention left to wonder which phrasing sounds fancier.

So how do you use it yourself? Install a skill too?

I paste it into the project's AGENTS.md:

Writing rules:
1. Write one action per sentence.
2. Keep a procedural sentence to 20 words or fewer.
3. Use the active voice and name the actor.
4. Use one name for one concept from start to finish.
5. Check the output against this list before you answer.

Output format:
1. Give the conclusion first.
2. Give numbered steps second.
3. Give risk notes last.

Does it work? Is it really that good?

I think it's fine. At least reading AI output takes less effort now.

I haven't run any benchmarks, though. It's all a feeling.

All right, I'll give it a try. Thanks, Sansi.

Anytime. Tell me how it goes.