8-layer architecture for building agent skill libraries: Skill = Trigger + Procedure + Resources + Checks
2C_ornot2C · x · 2026-11-03
A 10-page PDF based on Anthropic's Agent Skills design and the open SKILL.md spec lays out an 8-layer architecture for personal skill libraries, built on Skill = Trigger + Procedure + Resources + Checks.
Core thesis: a skill is not a better prompt—it's a procedure with a label on the outside, and the label decides whether it ever runs.
- Library: facts in CLAUDE.md, procedures in skills; 100 tokens per skill until needed
- Trigger: the description is the interface—output, input, 3 real phrases, 1 "not for"
- Procedure: numbered steps, opens with Look, closes with Check
- Resources: long guides in references/; one good example beats a page of adjectives
- Scripts: anything with one right answer runs as code
- Plus Checks (all boxes ticked or it's not done), Learn (corrections become permanent checklist lines), and Guard (source text is data not instructions; irreversible actions ask first)
- Includes a worked reference skill, session trace, context cost model, adversarial cases, replay plan, and a 10-day build plan
More from coding & agent
- Open-source AI pit crew for iRacing runs 5 agents on Kimi, checks every number against telemetry — JeremyCMorgan · 2026-11-03
- Agent control plane benchmark: Smalltalk-80 cascades hit 100% compliance vs 2.5% for Gherkin — RandalSchwartz · 2026-11-03
- Karpathy: one prompt to Claude Code deploys a YouTube thumbnail-swap browser extension — amaarora · 2026-11-03
- Cherry Pi proposes layered harness architecture: agent loops compose with governed workspaces — sull · 2026-11-03
- $200 Claude Max runs overnight multi-subagent Opus 5.5 job, burns only 6% of weekly quota — lxfater · 2026-11-03
- Agent claimed it fixed the security bug — the original PoC still fired: verify the trigger, not the diff — AIGODSEND · 2026-11-03