caveman-compress on rules/skills — evaluation

Decided: not worth applying /caveman-compress to ~/ai-config/AGENTS.md/CLAUDE.md or to SMTM skill/command files. Talbot confirmed 2026-07-26.

What was tested

1. AGENTS.md (329 lines, prose+rules doc): compression ran, words 4525→3681 (-19%), chars 31782→28155 (-11%) — well below caveman chat mode’s ~75%, since prose-memory compression must preserve headings/structure/exact code/URLs verbatim. Manual diff found no rule dilution. But the compressed output had a stray leaked line at the top of the file — meta-commentary from the compression LLM call (“Found it — dropped example, no file needed…”) that isn’t part of the source. The validator (checks only inline-code/URL preservation) didn’t catch it. → any real use would need a mandatory human diff-review pass every time, which erodes the efficiency case.

2. task-start.md (117 lines, procedural SMTM skill, heavy inline-code): compression failed validation twice (dropped literal instruction chunks containing inline-code fences); the script correctly gave up and restored the original untouched. No corruption, but zero benefit. Procedural skill files are dense with exact-literal strings that are the instruction (placeholder formats, checkbox syntax, command names) — the same density that makes them correctness-critical is what breaks the compressor.

Why not adopted

• Global rules gain (~15%) doesn’t justify the review burden + demonstrated defect risk on a file every agent session depends on for correct behavior.
• Changing AGENTS.md/CLAUDE.md requires CEO approval regardless (Core/AI/JOB_DESCRIPTION.md / Staff/AI-Director.md — A1 escalation).
• SMTM skills/commands: tool doesn’t work on them at all — not a close call.
• Talbot declined the narrower fallback (compress only the verbose “Discovered:” lesson-blocks inside AGENTS.md, skipping the inline-code-dense sections).

No live files were changed — testing was done entirely on scratchpad copies.