DEV Community

weiche chiu
weiche chiu

Posted on

Agent Skills Solved Distribution. Trust Is the Missing Layer.

Agent skills had their explosion. mattpocock/skills sits at ~269k stars, obra/superpowers at ~291k. Every coding agent, from Claude Code to OpenCode, now consumes SKILL.md files. Distribution is solved.

But here's what happened over the summer of 2026: the skills boom entered its governance phase. Three signals landed within weeks of each other, and together they point at the same missing layer.

Signal 1: NVIDIA put a number on the fear

NVIDIA/SkillSpector (~18k stars) is a scanner for agent skills, and it came with a quotable finding: ~26% of sampled skills contain vulnerabilities. Every SKILL.md install is unsigned code execution on a machine that holds real credentials — and until now, nobody had measured how bad it was.

Signal 2: Cloudflare shipped a playbook as a skill

cloudflare/security-audit-skill went from ~0.5k stars at its June launch to ~21.4k by late September — roughly 40x in three months. A vendor's security audit playbook, distributed not as SaaS but as an installable skill. "Playbook-as-skill" is now an official distribution channel.

Signal 3: Anthropic owns the onboarding path

anthropics/launch-your-agent guides you from an agent idea to a live managed agent. Read it alongside Vercel's Eve ("Next.js for agents") and the pattern is clear: platforms are splitting the agent stack into layers — framework, hosting, security. Each vendor claims one layer.

The gap: detection without a trust gate

Scanners tell you something is wrong. Nobody tells you:

  • Should I install this skill? (score)
  • Can it prove it's safe? (badge)
  • Can my team require that proof in PRs? (CI gate)
  • Can the obvious issues fix themselves? (minimal-permission rewrites)

That's the missing loop: lint → permission manifest → 0–100 score + badge → CI gate. The "eslint moment" for skills — neutral and vendor-independent. I built v0.1 as a Python CLI (sandbox trial-run and single-binary distribution are roadmap, not day-one promises).

What I'm doing about it

I started ai-gap-radar, a public biweekly research repo with a simple rule: hot repos are evidence, buildable gaps are conclusions. Each issue has a star-velocity radar, one-page briefs with verdicts, and opportunities/ — gaps concrete enough to ship, with competitor tables and 2-week MVP scopes.

The first issue tracks skills governance, and its first opportunity — a skill security linter — already has a working v0.1: https://github.com/williamlabdev/skill-security-linter. Day-one benchmarks: 0 false positives across 53 real skills (mattpocock/skills + obra/superpowers), 13/13 caught on known-bad patterns drawn from SkillSpector's own test suite, 30-second demo in the README. My validation rule is public: 500+ stars + 5 inbound "run this on my skills repo" issues → hosted scoreboard; otherwise it stays a CLI.

Corrections welcome — that's what a radar is for.

Top comments (0)