Agent Skill Forces LLMs to Write Docs in ASD-STE100 Simplified English
An MIT-licensed Agent Skill forces LLM output into ASD-STE100 Simplified Technical English, cutting rule violations 72.9% across six Claude models.
A new open-source Agent Skill forces LLM output into ASD-STE100 Simplified Technical English, the controlled-language standard aerospace has used since 1983 to keep maintenance instructions unambiguous. Packaged as a single MIT-licensed folder via the Agent Skills standard, it works across Claude Code, Cursor, VS Code Copilot, Codex, Gemini CLI and roughly 25 other harnesses, with a prompt-based fallback for tools lacking native skill support.
The skill enforces around 53 numbered STE rules: a 20-word limit per instruction, one-word-one-meaning vocabulary, active voice, a ban on hedging words like "should" and "may," and condition-before-command ordering. In benchmarks across six Claude models and eight writing tasks (96 generations), it cut STE rule violations by an average of 72.9% per 100 words, reduced output tokens on every model, and lowered mean sentence length from 11.2 to 9.7 words.
Beyond docs, the project ships adaptations for error messages, runbooks, incident reports, release notes, and even AGENTS.md system prompts, treating a system prompt as a procedure for a reader who cannot ask questions. It deliberately stays out of marketing and blog writing. The rules were derived from the primary Issue 9 (2025) ASD-STE100 specification, with each one tested against recorded baseline failures.