AI Pulse · Under the Radar

A developer named levelsio posted this week that he fixed Claude’s tendency to get wordy by telling it to always use ASD-STE100 Simplified Technical English. That’s a real standard, written in the 1980s for aircraft maintenance manuals. It limits vocabulary, bans metaphors, and keeps sentences short so technicians working on jet engines don’t misread instructions.

Within days, at least three GitHub repos packaged this as a reusable prompt skill. One claimed a 72.9% drop in what it called “AI writing violations”, though that number is self-reported and no one has verified it independently. The idea is simple. Claude, like most LLMs, drifts toward corporate fluff when you use it over time. Telling it to follow a strict technical writing standard seems to pull it back.

This matters if you ship anything with AI-written copy, especially product descriptions, help docs, or internal SOPs. The fix costs nothing to test. You add one line to Claude’s memory or system prompt and see if the output tightens up. If it does, you’ve saved editing time. If it doesn’t, you’ve lost five minutes.

The broader point is that people are solving real LLM problems by reaching outside AI entirely. ASD-STE100 wasn’t built for chatbots. It was built for clarity under pressure. Now it’s a verbosity filter. This is the kind of thing we build into an AI command centre, where you can set output rules once and apply them across every workflow that touches an LLM.

No one knows yet if this works at scale or if it breaks down in certain contexts. But it’s cheap, it’s testable, and it addresses a complaint we hear constantly. Claude writes too much. Now there’s a fix that doesn’t require fine-tuning or a new model.