You ask Claude a simple question about your code, and Claude replies with text like this:
I smoke-verified the config-guarded
materialization core, and the metadata-bearing
patches are spec-shaped.
The newest Claude models are strong, but their output has bad habits.
Claude invents project vocabulary, hedges, gives
three paragraphs where one sentence is
enough.
The good news: a fix exists. The setup takes two minutes. The fix comes from an unexpected place: aircraft maintenance manuals.
Lydia Hallie works on the Claude Code team at Anthropic. She shared a tip on X. Claude Code supports custom output styles. An output style is a file with permanent instructions. The instructions shape how Claude talks in every session:
drop instructions in ~/.claude/output-styles, then run /config → Output style
The key detail: an output style changes the system prompt itself.
A note in your CLAUDE.md file is different - that note is only one piece of context among many. An output style changes the voice at the source.
Andrew Carr is a developer and founder. He supplied the second half: the content of the file. A developer begged Claude for plain English. Andrew replied on X:
"only report to me in ASD-STE100 Simplified Technical English"
ASD-STE100 is a real standard that the European aerospace industry created in the 1980s. The goal: a mechanic must not misread a maintenance instruction. Many mechanics do not speak English as a first language. The standard permits approximately 900 words and 53 rules. Each word has one meaning. Each sentence gives one instruction, short and active.
The standard protects a tired mechanic from a misread torque value. The same rules stop an LLM from a self-invented dialect. Combine the standard with the "explain like I'm 5" format. The result is output that is short, plain, and correct.
You could do this with a simple prompt that works.
You must strictly communicate using a combination of the 'ELI5' format and the ASD-STE-100 (Simplified Technical English) language standard.
- Enforce a highly restricted, plain-English vocabulary.
- Do not invent, infer, or provide unmapped acronyms or domain-specific fluff.
- Lead directly with a bolded high-value takeaway or a single-sentence resolution.
- Break down walls of text into sequenced, scannable elements using concrete visual analogies.
The above prompt doesn't really work in the way it is forcing.
Give fuller prompts describing ASD-STE100 rules. Some can be found here:
Toppa's ASD-STE100 system prompt
Or, if you want to diverge from aircraft repair manuals style:
AmingBlg Simple English ruleset
There is an ISO standard for Plain English:
Some repositories that try to replicate it:
Output styles are Markdown files that Claude Code loads into its system prompt. You choose the scope.
mkdir -p ~/.claude/output-styles~/.claude/output-styles/eli5.md.claude/output-styles/eli5.md/config in Claude Code.The chat apps have no output-styles folder, but the same prompt works through two doors.
Every new conversation now starts with the plain-language rules already loaded.
I recommend this option to most people. Keep one plain-English workspace for code and technical work. Your other chats keep the default voice. The default voice helps in creative tasks.
What does a restricted vocabulary do to AI watermarks?
A watermark is a hidden signal in AI text. It shows that an AI wrote the text. The known methods, like Google DeepMind's SynthID-Text, push the model's word choices.
A sentence has many possible forms. The model prefers the form with the statistical mark. This method needs room. Normal English gives the model thousands of equal word choices in each paragraph. That room is the signal space.
ASD-STE100 removes exactly that room. The standard permits 900 words. Each word has one meaning. Each instruction has one obvious form. Few choices remain. So little signal space remains for the hidden mark.
I asked Claude Fable 5 to reword the above article in ASD-STE100 style, and you can see the result. I used Toppa's ASD-STE100 system prompt. I also asked it to give me two different article recommendation LinkedIn post offers.
A developer literally begged Claude: "Please just use plain English, I don't understand what you're saying."
Sound familiar?
The newest Claude models are brilliant at reasoning, but the output picked up bad habits - invented vocabulary, dense hedging, three paragraphs where one sentence would do.
The fix turned out to be 40 years old.
In the 1980s, the European aerospace industry built a writing standard called ASD-STE100 so that aircraft maintenance manuals could not be misread. Roughly 900 approved words. One meaning per word. One instruction per sentence.
Tell Claude to write in that standard, and the word salad stops.
I wrote up the full setup in my new article:
- Where the tip came from (an Anthropic engineer and a sharp developer on X)
- The exact prompt to paste
- Setup for Claude Code (global or per-repo) and the desktop app
Two minutes of work. Link in the comments.
Sometimes the best fix for new technology is an old industry that already solved the problem - and wrote it down.
#AI #ClaudeAI #DeveloperTools #PromptEngineering
I wrote this post in Simplified Technical English that aircraft manuals use.
The problem: Claude writes walls of text. It invents its own vocabulary. You read three paragraphs and learn one fact.
The fix comes from aerospace. European engineers built the ASD-STE100 standard in the 1980s. The standard permits about 900 words. Each word has one meaning, each sentence gives one instruction.
My new article shows the full setup:
- Claude Code, for one repo or for all repos
- The desktop app, for all chats or for one project
- The exact prompt, ready to paste
#AI #ClaudeAI #DeveloperTools #TechnicalWriting