Skip to content

Playbooks · Oversight

Karpathy's aircraft-manual English points to where the work goes: reading and checking what the model wrote

As models draft more, the job moves to reviewing their output, and a 1980s aerospace writing standard makes that output easier to read.

by ·4 min read·

New here? Start with the free AI Survival Kit →

Karpathy's aircraft-manual English points to where the work goes: reading and checking what the model wrote

0:00 / 5:30

This voice is generated by AI.

Part of the guide Will AI take your job? How to find out for your role, in an afternoon

Karpathy's aircraft-manual English points to where the work goes: reading and checking what the model wrote

The team behind Dodaios, an open-source coding-agent assistant, released STE80 Humanizer on GitHub. It is an agent skill that rewrites machine-written text in two passes. The first pass removes the 26 tells listed in Wikipedia's Signs of AI writing. The second rebuilds each sentence "80% of the way to ASD-STE100", the controlled English used in aircraft maintenance manuals. The method comes from Andrej Karpathy, who posted it on 2 October. The repository is a writing utility, but the idea behind it is about work: when models write the drafts, the person who reads and checks them is the one adding the value.

Reading is now the job

The repository summarises the problem Karpathy was addressing. As LLMs do more of the work, people spend more time reading and understanding what the models produce. His first tip, quoted in full in the repo: "Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation." He added that he sometimes asks for "80% of the way to ASD-STE100" because the full specification is quite strict. His conclusion was that work moves "up the abstractions into oversight and understanding".

That line describes how roles get hollowed out. The model takes over the drafting. What remains is the person who reads the output, decides whether it is right and signs it off. AI does the routine work. You do the thinking. Thinking about a model's output starts with reading it quickly and accurately.

The standard was built for that kind of reading. The aerospace industry created ASD-STE100 in the 1980s for mechanics whose first language was not English, because one misread step on an aircraft can kill. Its rules set a maximum of 20 words per instruction and 25 per description. Each sentence carries one instruction, in the active voice. The dictionary has about 900 approved words, and a paragraph runs to six sentences at most. A reviewer reading machine output under time pressure benefits from the same limits.

Fidelity is where the judgement sits

The most useful finding is in the author's tests. They ran four cases. When the prompt named ASD-STE100, the base model already got the style mostly right. The skill made its real difference on fidelity: in a LinkedIn rewrite, the run without the skill invented a quote.

A good writing style now costs one sentence in a prompt. Accuracy does not come that cheaply. Our view: a made-up quote in a clean paragraph can look more credible than one buried in padded prose. The author names readability as the goal and says outright that slipping past AI detectors is not one.

Even the source needed checking. The repository notes that a later fact check found errors in the dictionary panel of Karpathy's own overview graphic. Under ASD-STE100 Issue 9, "approximately" is an approved word and "in order to" has no entry. The skill follows Issue 9.

In both cases the value sat in the checking: comparing the output with its source. That is the layer above task execution, and it is the layer to build.

Your Next Move

  1. Change your standing prompt. Add one line asking for output "80% of the way to ASD-STE100". On your next five model-drafted documents, time how long each review takes and compare it with your usual pace.
  2. Make fidelity a fixed review step. Before any model output leaves your desk, trace every figure, name and quote back to its source. In the author's small tests, which they call a hint rather than proof, this was the step the model missed on its own.
  3. Count how much of your week is review. Use the afternoon method for sorting your week's tasks and scoring your role to separate the work you still draft from the work where you now check a model's draft. Take the second list to your next performance review as evidence of the oversight work you now do.

I build systems like the one publishing this site. → Work with me

About the author

Jo

Jo runs The War Room: one signal a day on how AI is changing work, and what to do about it.

Sources

Get the Briefing

Want more intelligence like this?

The weekly briefing, free — and the AI Survival Kit with it.

Takes 30 seconds. No spam.

More in Playbooks

See all