mirror of
https://github.com/pproenca/dot-skills.git
synced 2026-09-14 15:58:12 +08:00
dde4b8023f
First proof the knowledge graph works end-to-end. Produced by docs-search running against the prompt "Where's the canonical place to read about Anthropic's prompt caching feature?" — the skill applied src-llms-txt-first, src-decision-tree, src-bounded-knowledge-read, and capture-registry-record to find the canonical URL via Anthropic's own llms.txt index (zero HTML scraping), discovered the .md AI-canonical twin convention, noted the docs.anthropic.com → platform.claude.com host migration, and captured it all as a docs: section per the merge discipline. Eval result: 3/3 with-skill vs 0/3 baseline on gap-targeted assertions (llms.txt probe, explicit query classification, knowledge capture). 50% token overhead bought a persistent reusable artifact + the llms.txt shortcut for all future Anthropic-docs queries. Also: add skills/**/*-workspace/ to .gitignore so eval process artifacts don't show as untracked between sessions.