Consider: Simplified Technical English (STE) skill for reducing AI slop

[alycda]

Reference

https://github.com/woosal1337/blog/tree/main/videos/ep01-the-cure-for-ai-slop

What it is

Episode 01 of woosal1337's video series argues that giving an LLM a formal writing system beats banned-word lists or generic style rules for reducing "AI slop" (repetitive, hedgy, over-qualified AI prose). It adapts ASD-STE100 (Simplified Technical English — a controlled-language standard from 1986 aircraft maintenance manuals) into an LLM-usable skill, and backs the claim with cross-model measurements:

• ste-writing-skill.md — a distilled ASD-STE100 skill, with strict/flexible modes
• ste-lint.py — a heuristic linter that scores violations per 100 words
• Reported results: 74% violation reduction on Claude Sonnet, 50% on GPT-4o — vs. only 3–40% from banned-word-list approaches

Why this might be relevant here

This repo already carries agent-instruction content that shapes how Claude writes in this codebase:

• tools/agents/ (AGENTS.md overlay, company-values.md, personal-constitution*.md, code-critic.md, constitution-critic.md, rubrics/, skills/)
• tools/claude/rules
• lib/skills-sh.nix — the existing pattern for pulling in skills.sh skills declaratively via nix-skills

Worth a look to see whether an STE-style writing skill (or its linting approach) is worth adopting here — either as a tools/agents/skills/ addition, wired through lib/skills-sh.nix if it's indexed by nix-skills, or just as a technique to fold into the existing critic/rubric docs.

To consider

• Read ste-writing-skill.md and ste-lint.py in the linked repo
• Check whether woosal1337/blog (or the skill specifically) is indexed in nix-skills / skills.sh
• Decide: adopt as-is, adapt into an existing critic doc, or pass
• If adopted, document the decision per the "Meta: Updating This Document" section of CLAUDE.md