Karpathy's trick: use a 40-year-old aviation writing standard to make LLMs clearer
机器之心 · wechat · 2026-10-03
Karpathy shared techniques for making LLM output easier to understand, surprisingly based on the 40-year-old aerospace controlled-English standard ASD-STE100: keep procedural instructions under 20 words, one action per sentence, active voice, and one consistent name per concept. He suggests asking for "80% of ASD-STE100" rather than the full rigid spec. Beyond text, he advocates having models generate diagrams, interactive HTML pages (e.g., a slider demoing learning rate), and fully custom explainer videos in 3Blue1Brown style with ElevenLabs narration — what he calls large, custom, discardable software artifacts. The community tested it: gains are most visible on long, complex outputs; one workflow has the model audit your last ten chats against the standard and write the winning rules into your user-level AGENTS.md; and the spec has already been packaged as an open-source Claude Skill with Strict and STE-flavored modes for tool descriptions, agent instructions and system prompts.
Related event: Karpathy Shares Four Techniques for Understanding LLM Output(7 posts)→
More from coding & agent
- Fewer, well-designed agent harnesses beat sprawling agent fleets, argues vixsheikh — annetgriffin · 2026-10-26
- ConformWeb: LLM-generated web apps pass only 11.7% of behavioral contract checks — algo_diver · 2026-10-26
- Agent memory explained in 12 minutes: working, procedural, semantic and episodic — _AustinCalvert_ · 2026-10-26
- Grok Bot v0.75.5 leak touts stored cards, email forwarding, 3x automation limits — mark_k · 2026-10-26
- Every time my agent breaks my UI, I add a rule to its skill — 10 design laws — _AustinCalvert_ · 2026-10-26
- Turn nouns into verbs: the 'Range' trick to unlock cross-domain AI output — petesena · 2026-10-26