A source can provide evidence, never a safety guarantee. Review provenance, the pinned package and dependencies, required permissions, test evidence, ownership, monitoring, and removal before adoption.
Reviewed 31 August 2026. Anthropic's current Agent Skills overview defines a skill as a filesystem package that may include executable code. Model-assisted review can help find questions; it cannot certify a package or replace manual code review.
Start with provenance, then apply the same technical review to every source. Popularity, branding, and public visibility reduce some uncertainty but do not prevent compromise or unsafe updates.
Confirm the exact publisher through current official documentation, then record the release or commit and integrity value. Publisher identity does not eliminate bugs, compromise, or environment mismatch.
Treat as: strongest provenance; still inspect and test.
History, issues, ownership, and diffs provide evidence. Stars and contributor counts are social signals, not controls; verify tags, dependencies, release artifacts, and maintainer changes.
Treat as: pin, inspect every file, and test.
Identity, integrity, and update history may be absent. Do not install it to inspect it and do not upload confidential contents to an unapproved reviewer.
Treat as: quarantine; obtain an authoritative source or reject.
Do not rely on a visual badge or a changing install count. Keep a small review record so another person can reproduce the decision and identify what changed later.
Approval applies to the exact bytes and configuration reviewed—not to a package name forever. Pin versions or hashes where the distribution mechanism allows it, review every update, and remove approval when ownership, contents, permissions, or dependencies change unexpectedly.
After a human inventories and reviews the complete package, a model can surface missed questions. Use an approved environment and sanitized content; never paste credentials, customer data, or proprietary code into an unapproved service.
Verify each finding in the source. A clean model response is not due diligence: models can overlook behavior, misunderstand code, or hallucinate. Unknown binaries, missing dependencies, or code your reviewer cannot assess remain unresolved.
List every file, type, size, link, generated artifact, and archive member. Keep it outside project and personal skill directories; do not execute preview helpers.
Trace the maintainer through an authoritative channel. Record an immutable revision and hash or signature where available; inspect release and ownership changes.
Map inputs, outputs, scripts, subprocesses, tools, file paths, endpoints, data fields, credentials, destructive actions, and failure behavior. Get qualified review for unfamiliar code.
Scan dependencies and secrets, inspect diffs, and run the prompt above only with sanitized contents in an approved service. Resolve findings in source; never auto-approve a model verdict.
Use an isolated environment without secrets or production data. Require explicit approval for writes, messages, purchases, deployments, or other external effects. Capture logs, failures, and rollback; assign an owner and re-review date.
Re-review every update and periodically reassess unchanged packages as runtimes, dependencies, permissions, and threats evolve. If behavior deviates, disable or remove the skill, preserve logs, rotate any possibly exposed credentials, and follow the documented rollback or incident process.