An open source tool on GitHub forces LLMs to write in ASD-STE100, the controlled language that aerospace has used since 1983 because a tired mechanic shouldn’t misread an instruction. AminBlg’s SimpleEnglish repo hit 333 points on Hacker News in 24 hours.

What it does. Instead of asking the model to write clearly and hoping for the best, you load it with a skill that applies Simplified Technical English rules: restricted vocabulary, short sentences, one verb per instruction, no ambiguity. The before/after examples in the repo transform typical output full of “seamlessly” and “leveraging” into dry technical manual instructions.

Why it matters to you. If you use agents to produce documentation or instructions for non-specialist audiences, AI slop is a daily problem. Here the approach is to constrain writing at the source, not fix it after. The numbers in the repo (96 runs across 6 Claude models for 8 tasks, data from July 30, 2026) show STE rule violations down 72.9% on every model, average sentence length from 11.2 to 9.7 words, and the word “seamlessly” eliminated.

The skill follows the Agent Skills standard: it runs in Claude Code, Cursor, Copilot, Codex, Gemini CLI and other environments. The repo provides installation instructions and a version to paste into your system prompt for those without native support. MIT license, no dependencies.

In detail

ASD-STE100 is a standard formalized in 1986 (with precedents from 1979) by the European Association of Aerospace Industries, built for a real need: aerospace maintenance manuals are read by technicians worldwide, often non-native English speakers, often under pressure or tired. An ambiguous sentence in those conditions costs lives. The standard solves it at the root with a vocabulary of roughly 900 approved words, restrictive grammar rules (one sentence, one instruction; no passive voice when it creates ambiguity; one word per concept, no synonyms) and a fixed writing structure.

What changes with an LLM. Models produce fluent text but follow the style of their training data, which is full of conversational English, marketing, and the verbosity we call AI slop: “leveraging”, “seamlessly”, “robust”, ceremonial apologies, nested clauses. Asking the model to “write clearly” isn’t enough, because the model has no formal criterion for what “clear” means. The skill injects that criterion: STE rules become part of the agent’s system prompt or context, and the model follows them like any style instruction.

How it’s measured. The repo reports 96 comparisons: 6 Claude family variants, 8 technical writing tasks (README, error messages, incident reports, release notes), each with and without the skill. The declared result is a 72.9% reduction in STE violations per 100 words, with improvements across all tested models. Average sentence length drops from 11.2 to 9.7 words, and output tokens decrease, because controlled language is also shorter.

Where to be careful. The numbers come from the repo itself, not independent evaluation. The 6 models tested are all Claude: missing are GPT, Gemini, open models. The before/after examples in the README are declared unretouched, but they’re selected: we don’t know what percentage of output actually improves and how much still needs manual fixing. ASD-STE100 is a standard for technical English: those writing documentation in Italian or other languages can draw inspiration from the rules but can’t apply them mechanically.

The broader idea. The pattern is interesting beyond STE: instead of fixing AI output after it’s produced, you constrain it while it produces with a set of formal rules. It’s the same logic as the technical jargon translator on this site, where you declare your audience and glossary before letting the model work. The difference is that here the rules are a stable formalism, proven over decades, not a list written on the spot.