Ask for a Clearer Format
When model output is hard to read, change the format before you change the question: controlled English, then a diagram, then a web page, then a video.
The prompt
"Ask your LLM to explain something in ASD-STE100" (Andrej Karpathy)
ASD-STE100 is Simplified Technical English, a controlled language written for aircraft maintenance manuals. Karpathy's ladder goes from there to diagrams, to answers "in HTML" as interactive pages, to custom explainer videos ("a 3b1b style video explainer on X"), and he sometimes softens the first rung to "80% of the way to ASD-STE100". Source: X/@karpathy, 2026-10-02 @daniel_mac8's version adds Zinsser's four principles of good writing: simplicity, brevity, clarity and humanity. Source: X/@daniel_mac8, 2026-08-09
Why it works
A model's default prose is long, hedged and padded, because that is what its training rewarded. A named standard replaces vague taste ("be clearer") with rules the model already knows in detail: STE limits sentence length, bans most synonyms and requires one meaning per word. The model can follow a standard far more reliably than an adjective.
The rest of the ladder works for a different reason. As agents do more of the work, people spend more time understanding what was done. Code and intelligence are now cheap enough to build a throwaway diagram, page or video for one explanation, which was never worth it before.
When to use it
- Reading an agent's report, plan or explanation of unfamiliar code.
- STE for technical answers; it is too stiff for writing meant to sound like a person.
- A diagram for structure and flow, HTML for anything with state to explore, a video for a concept you want to remember.
Which skill runs it
No single skill. diagram-to-html and visual-plan produce the diagram rung, an artifact page the HTML rung, and hyperframes the video rung. Prose Kevin publishes goes through humanizer instead.
A writing kit around it
Ruben Hassid builds a kit around the STE prompt: list the AI-writing tells (the "that's not X, that's Y" turn, two clipped bits stuck together, self-applause, warm-up phrases, constant triplets, ranges, recap endings); use "Use ASD-STE100" for guides and emails but not for poems or posts; add Zinsser's four principles to restore the humanity STE drops; keep a file of forbidden patterns and check every draft against it; and start a fresh chat when a long one starts repeating itself. Source: X/@rubenhassid, 2026-08-13 Kevin's humanizer skill works from the same kind of list.
Other named styles
Any well-known style guide works the same way. "Make it like Google dev docs style. More dead prose. No aphorisms, no flourishes." Source: X/@aaronvi, 2026-08-12 One user put STE in a global CLAUDE.md with its five core rules (approved words, one word per idea, short sentences, active voice, short paragraphs) and reports every agent became more concise. Source: X/@topmass, 2026-08-03, image reviewed
A follow-up from the person who passed the STE tip on: after adding it to agent rules, the fake drama ("this is the smoking gun") and waffling technical prose were gone, so the rule stayed. Source: X/@DanielLockyer, 2026-08-06
Two skills package the writing side: blader's humanizer, built from Wikipedia's "Signs of AI writing" page, and stop-slop, which scores a draft on directness, rhythm, trust, authenticity and density and rewrites anything under 35 of 50. Kevin's humanizer skill folds in both. Source: X/@adrknapp, 2026-08-03
Answer-first replies
The i-have-adhd skill reshapes replies the same way STE reshapes sentences. Its before-and-after: a paragraph that opens with "Great question!" and wanders, against the action first, then numbered steps with file and line, then one next step. Source: X/@denysdovhan, 2026-07-20, screenshot reviewed; Source: X/@denysdovhan, 2026-07-20; Source: X/@MiguelriosEN, 2026-07-20
Put rules like this in AGENTS.md, not only in a skill, because models do not always load a matching skill; a standing instruction always applies. Source: X/@kaushikgopal, 2026-07-21
The post that spread i-have-adhd: "it's made my claude replies so good." Source: X/@jjacky, 2026-07-19; Source: X/@jjacky, 2026-07-19
Simon Willison's LLM cliché highlighter marks the stock phrases in a draft, a quick check after asking for plainer prose. Source: X/@simonw, 2026-07-18
The ADHD-style reply skill spread further as an open-source Claude Code skill. Source: X/@Saboo_Shubham_, 2026-07-20
Timeline
- 2026-11-03 | Page created from Karpathy's output-format ladder and the STE writing tip. Source: X bookmarks, 2026-08-09 and 2026-10-02