Borrowing aviation's controlled English (ASD-STE100) to strip AI fluff and fake completions
sujingshen · x · 2026-10-05
The author proposes using ASD-STE100, the aviation industry's controlled English for non-native maintenance crews, as a style guide for prompting LLMs: one word one meaning, one sentence one idea, active voice only, and never mark unfinished work as done. Since models saw tons of such manuals in training, this style reliably eliminates filler, tautology, and fake completion claims.
The author distilled it into a short markdown rule set:
- Controlled-English style constraints
- Plan before acting
- Deliverables limited to four sections: completed, verification results (with exit codes), not completed, specific blockers
- Explicitly excludes casual chat and brainstorming, otherwise the model turns conversations into work orders
The result isn't a smarter model — it's that errors surface instead of hiding behind three screens of fluff. Rules are in the reply; the author argues this beats adding "you are a professional assistant."
Related event: Using aviation maintenance controlled English to discipline AI output(2 posts)→
More from coding & agent
- Cohere launches North 2 with 15+ features for enterprise-grade agentic AI — cohere · 2026-10-05
- Stack Internal hits GA: free workspaces bring verified knowledge to enterprise AI agents — pchandrasekar · 2026-10-05
- LangChain Brings Interrupt Agent Conference to London on October 13, 2026 — LangChain · 2026-10-05
- Claude Code VS Code extension bug: session history ignores CLAUDE_CONFIG_DIR when a process wrapper is set — dfinke · 2026-10-05
- Otis: an open-source AI agent that auto-sets up llama.cpp and unifies local and hosted inference — Objective-Pair8231 · 2026-10-05
- Alibaba's FlyAI Skill Lets Claude Code Search Real-Time Flights and Hotels — tom_doerr · 2026-10-05