Making Claude write like a maintenance manual

I gave Claude a new rule: write like somebody could die if the sentence is misunderstood.

That sounds dramatic until you look at where the rule came from. ASD-STE100 Simplified Technical English was built for aerospace maintenance documentation. Its job is not to make prose pleasant. Its job is to stop a mechanic, possibly working in a second language, from reading one sentence in two ways.

I found the idea through a tweet by levelsio. The pitch was simple: make an AI assistant use controlled English and see whether its answers improve.

So I built a /ste skill for Claude Code.

What the skill changes#

The skill applies a practical subset of ASD-STE100. Instructions get 20 words per sentence; descriptions get 25. Claude must use active voice, simple tenses and one instruction per sentence. Conditions come first:

If the light comes on, push the button.

It bans contractions, idioms, long noun clusters and most -ing forms. “Use the tool for cutting” becomes “Use the tool to cut.”

The vocabulary rules matter too. Pick one plain word and keep using it. “Remove,” “delete” and “eliminate” may look interchangeable. A reader can reasonably wonder whether they describe three operations.

Models seem almost embarrassed to repeat a noun. This is considerate of them, but not especially useful.

The skill protects anything that must stay exact. It does not rewrite code, identifiers, paths, commands, logs or technical names. AVContentKeySession stays AVContentKeySession. Renaming it “the key thing” would improve nothing.

More machine, not less#

STE makes Claude sound more like a machine.

That is expected. A maintenance manual should be boring and predictable. The same word must mean the same thing every time. Nobody needs voice or nuance in an instruction to disconnect a high-voltage cable.

This is why I did not turn it on for everything. It suits procedures, specifications and technical explanations. It is less useful for email, blog posts and tickets. If every message resembles an aircraft service bulletin, people may stop reading before safety improves.

The skill has two modes. /ste switches the conversation into the controlled register. /ste <text> rewrites one piece of text. The second is probably more useful. I can write normally, then apply STE where precision matters.

This is not really ASD-STE100 yet#

The standard has roughly 65 writing rules and a dictionary of about 900 approved words. Each approved word has one meaning and one part of speech.

My skill has the rules. It does not have the dictionary.

The official dictionary is a free PDF after registration, but I have not bundled it. The skill falls back to common words without sacrificing precise technical terms.

The skill is based on STE, not a conforming implementation. The controlled dictionary is what turns a writing style into a controlled language.

A skill or an output style?#

Claude Code can load an output style for a project. That looks like the natural home for STE.

There is a trap. Output styles default to keep-coding-instructions: false. This removes Claude Code’s normal software-engineering instructions about scope, comments and verification. It makes sense for a style that turns Claude into a tutor. It is a bold choice for a coding project.

An STE style needs this:

---
name: ste
description: ASD-STE100 Simplified Technical English
keep-coding-instructions: true
---

Subagents create another problem. They do not inherit the main agent’s output style. The conversation could use STE while a delegated plan or pull-request body returns in Claude’s usual register.

That inconsistency is why STE remains a skill for now. I can invoke it when needed and collect real examples before making it a project default.

What I am trying to fix#

Claude’s default technical prose is often dense and needlessly complicated. It stacks qualifications, abstract nouns and long sentences until a simple answer looks sophisticated. The result can sound impressive on first read while making the actual point harder to find.

I regularly have to unpack a response that could have been three plain sentences, which is not a great use of either intelligence involved. STE gives me a blunt tool against that habit. The short sentences and fixed vocabulary leave Claude fewer places to hide unnecessary complexity.

I do not expect every answer to read like a maintenance manual. I want the useful part of that discipline: say exactly what happens, use one name for each thing, and stop when the explanation is complete.