The largest tell for shifting culture in AI tools is often X, viral GitHub repos, and Hacker News. Lately, skills like "I have ADHD" and instructions to use ASD-STE100 Simplified Technical English have become popular. While the appeal is understandable—many dislike LLM verbosity—humanising the output at the model stage is the wrong abstraction. These instructions become part of the work, compressing output into a lower-bandwidth format that is lossy. For instance, ASD-STE100 was designed for human technical writers, not agents. The raw state is usually the most information-dense representation.

This issue intensifies when agents communicate: a subagent investigates a bug, turns findings into a human-readable summary, then the parent agent summarizes again. Instead of "Most tests passed, although there was one issue worth looking into," we should get "5/6 PASS/FAIL: test_cache_invalidation CAUSE: stale key survives restart REPRO: tests/cache_test.py:184". Humanisation hides failure. Agents fail in ugly, useful ways—conflicting evidence, stack traces, uncertain assumptions—but human prose smooths them into "There are a few considerations here."

Other systems work differently: databases don't store in display format, compilers don't make IR pleasant. Keep high-fidelity representation until the human boundary. Let agents exchange schemas, diffs, exact errors, confidence, provenance. Then compress. Skills like "Talk to me like I have ADHD" make sense as a renderer, not operating instruction. The viral repos are a bug report, not the end state.