skillspec

The pitch here is narrow and defensible: agent skills are prose, and prose can be skipped, reframed or forgotten. SkillSpec's answer is to stop trusting the prose and start checking it. Import turns a SKILL.md into routes, rules, forbids and regression tests; Doctor scores the risk that an agent won't follow it; boundary maps what it could reach if it does; alignment replays the decision trace afterwards and returns aligned, partial or unproven. That last verdict matters — the tool refuses to collapse uncertainty into a pass/fail badge. Strengths. Boundary stands alone, which is the smartest design decision in the package: you point it at a local folder or a public git URL on any host, it needs no skill.spec.yml, and it executes nothing. The risk model is scope-aware rather than keyword-scared, so a skill that only touches its own directory gets a low ranking while one reaching ~/.aws/credentials or an external host gets flagged. The browser flow — paste a public GitHub skill URL, get a follow-through and security read, nothing executed — is the lowest-friction way to vet something before you clone it. The vendor's refusal to count documentation examples as live effects is a small detail that says a lot about how the analyzer was built. And the research framing is unusually disciplined, citing Snyk ToxicSkills 2026, a 40,285-skill dataset with 46% name collisions, OWASP's malicious-skills entry, and the governance-decay work on context compaction erasing safety constraints. Weaknesses. The vendor is explicit that SkillSpec does not enforce the boundary — a harness, hook or network policy has to do that, so the compiled policy is only as strong as the layer you wire it into. Analysis is static and nothing is executed, so runtime behaviour that only appears on execution is out of scope; alignment can legitimately return unproven, which will frustrate anyone who wants a clean green check. It is an add-on to an existing agent setup, not a replacement runtime, and the whole flow assumes terminal comfort. Contract overhead on a throwaway two-line skill is real work for no payoff. Where it fits. Platform teams running large skill catalogs, security-conscious agent developers who vet before installing, compliance-heavy environments that need replayable traces, and Claude or Codex users who want install-time gating via skillspec pull and a managed pre-tool hook. Where it doesn't: no-code users expecting a GUI, and small experiments where the contract layer costs more than it saves. The honest positioning is that SkillSpec is a verification and audit layer, not a runtime — buy it for the boundary map and the trace, not for enforcement.