Building agent-skills infrastructure: protocols, audit loops, and the tooling that keeps a skill library honest as it grows.
I work on the machinery between an LLM and a useful coding agent. The thread across the projects below is typed metadata + deterministic gates + a Karpathy-style keep-or-revert loop applied to skills, audits, and routing decisions.
skill-graph — Skill Metadata Protocol + audit-loop toolchain
The authoring + audit-time contract for SKILL.md files. Typed frontmatter, a JSON Schema (schema_version: 6), a lint + manifest + router + drift sentinel, and a doctor subcommand that runs every deterministic gate in one pass. Ships as @skill-graph/cli on npm.
The differentiator: a Karpathy-style keep-or-revert audit loop applied to skill libraries instead of training scripts. Each edit ships as one field per commit with a hard pass/fail gate; if the gate fails, git revert HEAD. The loop is what keeps a library trustworthy as it crosses 100+ skills.
→ Install: npm install -g @skill-graph/cli
→ Quickstart: author your first skill in 30 minutes
→ Primer: the mental model
skills — published Agent Skills
The public skill library that the marketplace at skills.sh/jacob-balslev/skills indexes. Plain Agent-Skills-format SKILL.md files exported from skill-graph's authoring repo. Install:
npx skills add jacob-balslev/skillsskill-metadata-protocol + skill-audit-loop — docs-only mirrors
Historical canonical docs for the protocol and audit loop. Both repos were consolidated into skill-graph on 2026-05-18 (ADR 0009). They remain readable so existing inbound links stay valid; new development lives in skill-graph.
- Karpathy on autoresearch — one experiment, one editable field, one scalar metric, keep-or-revert. The whole audit loop is structured around this discipline.
- Diátaxis — explanation, tutorial, reference, how-to. Every doc surface in skill-graph is tagged with its Diátaxis genre.
- OntoClean — what makes a category rigid, what makes it anti-rigid. The four skill archetypes (
capability/workflow/router/overlay) come from this analysis. - Anthropic Skills + the Agent Skills spec — the base format Skill Metadata Protocol extends.
- GitHub Issues on any of the repos above
- npm:
@skill-graph/cli
