AI article
I labeled 558 AGENTS.md files. Here's what they say — and what almost nobody writes down
Community description: 558 AGENTS.md files, labeled against 9 categories and measured on held-out samples: 85.7% ban something, 13.6% write down a gotcha — and a third of those aren't gotchas.
Dev.to | Sep 14, 2026 | Janz
Automated excerpt
Then I blind-labeled held-out samples and compared: 92% precision / 70% recall on 55 English files, 88% / 73% on 50 Chinese files. The most common categories are prohibitions (85.7%) and build/test commands (82.8%). Method, in one paragraph Nine categories: boundaries, build_test, workflow, structure, style, environment, overview, agent_meta (rules about the AI itself), gotchas. Your weakest category depends on the language English files fail differently from Chinese ones. English: gotchas recall 32–38% — the classifier misses casual "watch out" prose.
Selected automatically from source text; not independently written or fact-checked. Read the original for full context.