skillkit

Tooling for the agent-skill economy: lint, package, verify, and audit SKILL.md skills.

skillkit lint ./my-skill        # strict frontmatter + structure lint
skillkit pack ./my-skill        # deterministic archive -> content-addressed store
skillkit verify ./my-skill --hash <blake3> --store ~/.local/share/skillkit/store
skillkit audit owner/repo       # engagement-integrity signals (fork-ratio heuristics)
skillkit audit anthropics/claude-plugins-official --batch
                                # audit every repo-plugin in a marketplace index

Batch mode / trust reports

audit --batch parses a marketplace's .claude-plugin/marketplace.json, audits every repo-sourced plugin, and prints flagged-first evidence. The first run against Anthropic's official directory is checked in at reports/claude-plugins-official-2026-10-19.txt: 227 of 315 entries audited — 1 EXTREME (adobe/skills, 41.8% fork ratio), 9 elevated (Salesforce, Oracle, GCP, Databricks, Microsoft — the fork-to-customize vendor pattern).

Why

Agent skills are becoming the Dockerfiles of AI agents — and the ecosystem has formats and marketplaces but no tooling layer:

• Lint with unknown-field denial: the frontmatter grammar is tiny and anything else is a finding, because a format that accepts everything validates nothing.
• Package deterministically: same directory ⇒ same BLAKE3 hash ⇒ free deduplication, hash-verified updates.
• Audit engagement integrity: fork-ratio verdicts and other heuristics from signalkit-integrity, because trending metrics get farmed (forks are cheap; 46% fork ratios exist in the wild).

Layered on the estate

Apache-2.0. Part of the WyattAu toolkit estate.