PART 07 · PHASE 3

Validate, release & maintain.

A release is an approval for exact package bytes, dependencies, permissions and environment—not a permanent safety badge. Distribute only through a surface you have verified and keep a removal and incident path.

20–30 MinGovernance

Reviewed 31 August 2026. The earlier “official marketplace,” verified badge and one-click install claims were removed because they were not supported by Anthropic's documentation. Product surfaces change; verify the current Agent Skills overview before release.

Current surfaces have different boundaries

Claude Code

Filesystem skills

Project skills live under .claude/skills/; personal skills under ~/.claude/skills/. A project repository can distribute reviewed skill files, but repository visibility is not a safety review.

Runtime: host tools and network subject to effective permissions
claude.ai

User-uploaded custom skills

Eligible users can upload custom skills through current product settings when code execution is enabled. Verify availability, sharing scope, retention and admin controls for your plan.

Runtime: depends on product/admin configuration
Claude API

Workspace skill versions

Custom skills are uploaded and versioned through the current Skills API and used in a code-execution container. This environment has no external network access or runtime package installation.

Runtime: constrained container; workspace governance required

Release evidence

1
Identity — authoritative maintainer, owner, license, contact and support horizon.
2
Integrity — immutable revision and hash/signature where available; manifest every file and dependency.
3
Reach — instructions, scripts, subprocesses, tools, files, network endpoints, transmitted fields and credentials.
4
Evaluation — synthetic representative, edge, adversarial and regression cases with a written rubric and retained failures.
5
Decision — approver, scope, environment, minimum permissions, expiry, monitoring, stop condition, removal and rollback.

Safe release workflow

  1. Freeze a candidate version; inspect every file, script and dependency manually. Model review is only a fallible second pass.
  2. Test in an isolated environment with synthetic or rights-cleared fixtures, no secrets, no client data and the minimum possible permissions.
  3. Verify expected behavior and attempts to read unrelated files, invoke tools, access the network, hide actions, bypass approval or exfiltrate data.
  4. Document installation and removal without download-and-execute shortcuts. Pin the reviewed version and state the supported surface.
  5. Require explicit approval for consequential actions; place real authorization in runtime and application controls, not instructions.
  6. Release to a small audience first, monitor tool/file/network activity and user reports, and preserve logs needed for investigation.
  7. Re-review every update and dependency change. Revoke approval, remove the package and rotate exposed credentials when trust changes.

Repository checklist

A public repository can improve inspectability but does not certify safety. Include:

Never approve a moving branch, a package name, a reputation badge or an AI-generated “looks safe” summary. Approval covers the reviewed bytes and configuration only. External URLs and dependencies can change after review; avoid them where possible and pin/verify them where required.

Maintenance is part of the skill

Schedule re-review for ownership, source, scripts, dependencies, permissions, platform behavior and data policy. Retire a skill when its owner, evidence or environment is no longer trustworthy. Safe deletion also includes removing cached copies, revoking permissions and rotating credentials if exposure is possible.

You finished the reviewed playbook.

You can now treat a skill as what it is: a filesystem software package whose reach depends on its runtime and permissions. The durable workflow is provenance → every-file inspection → least privilege → isolated evaluation → exact approval → monitored release → re-review and removal.

The output of this course is not a badge. It is an evidence record another reviewer can reproduce.

← Part 6 — Evaluate behaviorBack to playbook →