SimpleEnglish agent skill cuts Claude STE violations 72.9%

For anyone who has tried to prompt an LLM into writing plainly, the standard move is a variation on 'write clearly, avoid jargon.' It mostly does not work. A GitHub project called SimpleEnglish, released by AminBlg, tries a different tack: instead of asking a model to be clearer, it forces the model to write in ASD-STE100 Simplified Technical English, a controlled language aerospace has used since 1983 so a tired mechanic cannot misread an instruction. The current edition, published January 2025, is 53 numbered rules plus a dictionary of about 900 approved words.

The measured effect is what makes this worth a look. Across six Claude models and eight writing tasks, 96 generations in total, the skill produced 72.9% fewer STE violations per 100 words than the baseline, and output tokens went down on every model. The per-model spread is uneven: claude-opus-4-7 and claude-opus-4-6 each dropped 82%, claude-sonnet-5 landed at 80%, and claude-opus-4-8 came in lowest at 41%. The comparison uses 'a deterministic regex linter, same rules for both conditions,' which is why the numbers are stable but also what defines the ceiling of the claim.

The intuition the author leans on is that the aerospace rulebook happens to be a near-perfect negative of every AI writing tell. Max 20 words per instruction. Simple tenses only, no '-ing' forms. Active voice. One word per meaning, so 'start' is approved while 'begin' and 'commence' are not. Conditions before commands. An LLM constrained this way cannot produce the softening throat-clearing and hedged sub-clauses that read as bot output, because the grammar the constraint permits does not contain them.

The honest caveats are worth stating. What the linter measures is rule violations, which is a proxy for readability, not readability itself. The skill is a single-developer, MIT-licensed folder with no dependencies, and it explicitly refuses application to 'marketing copy, blog voice, brand writing.' The project also affirms it is unofficial and unendorsed by ASD or STEMG. What the repo does not give you is any independent human evaluation of the output, or a like-for-like comparison against non-Claude models beyond the claim that the same paste-in system prompt works in ChatGPT and Gemini.

The most interesting readers are probably technical-writing teams and anyone shipping API or hardware documentation where ambiguity has a cost measured in support tickets. If a portable set of rules can push six different frontier models toward the same registered voice, the case for treating LLM docs as a first-draft engine rather than a liability starts to look less speculative.