skill-family-curationlisted

<!-- S305 gate run (private research repo, quarantine gate-0 audit): added disable-model-invocation per the topology rule in AGENTS.md. This card prescribes physical deletion of SKILL.md files and atomic commits, which makes it a PROCEDURE with side effects, not a model-pulled ability. It was authored without the flag. Cleared Gate 0; standing is CANT_TELL_YET pending a screen. Not promoted. --> # Skill Family Curation ## Problem A skill library grows over time. Each /claudeception extraction adds one or more skills to `_quarantine/`. Manual §1.5 reviews promote some to active. Without curation, three failure modes compound: 1. **Skill bloat**: many skills with overlapping triggers. Semantic search returns 4-5 candidates for one query; the agent doesn't know which to invoke; the user mental-model breaks. 2. **Doc rot**: a CONSOLIDATED-into-parent skill leaves stale child skills sitting in the directory. The agent's context window gets polluted; the agent gets confused over which one to apply. 3. **Theater**: the LLM, asked "is this a family?", will generate plausible English justifications regardless of whether the consolidation is actually warranted. LLMs are optimized to reward plausible-sounding answers; the easy path is to confidently assert a family exists and propose a manufactured parent. The naive response — "have the LLM periodically scan skills and propose consolidations" — fails on all three. It accelerates bloat (more skills, more consoli