Borrowing ASD-STE100 controlled English to discipline LLM outputs and delivery reports
sujingshen · x · 2026-10-05
Instead of begging models with "please don't hallucinate," the author applies rules from ASD-STE100 — the controlled English used in aviation maintenance manuals — to constrain model outputs. Because models have seen countless such manuals in training, naming the style measurably cuts filler, synonym churn, and false completion claims.
The distilled constraint file covers:
- Style (80% ASD-STE100): one fact per sentence, active voice, fixed naming, no pleasantries or unverifiable adjectives; write "uncertain" explicitly when unsure.
- Interaction order: confirm task boundary and done criteria first; ask only blocking questions; get approval before side-effectful actions; revise plans per veto reason; stop and report new blockers.
- Evidence rules: never mark unrun checks as passed; citations must be locatable (paths, commands, exit codes); failures must include the actual error.
- Delivery format: task replies use exactly four sections — done / verification / not done / specific blockers.
The point isn't a smarter model, but errors surfacing themselves. Casual and brainstorming tasks are out of scope.
Related event: Using aviation maintenance controlled English to discipline AI output(2 posts)→
More from coding & agent
- Agent buys itself a Mac Mini: price comparison, reservation and setup recap — Baconbrix · 2026-10-05
- Cloudflare ships 46 announcements in Birthday Week, bets on agentic internet economy — threepointone · 2026-10-05
- Hugging Face Hub adds RL Environments: envs as dataset repos, no registries — ben_burtenshaw · 2026-10-05
- Reddit: What Does Your Real-World Coding Agent Sandbox Setup Actually Look Like? — radim11 · 2026-10-05
- Microsoft's ThinkingBox: grade agents by database changes, not their words — SergioPaniego · 2026-10-05
- Prompt bloat in a RAG app: building evals from GitHub issues and accepted PRs — Al_Grigor · 2026-10-05