GitHub shares reusable agent skills with independent verification

GitHub’s November 12 guide uses Proof Agent and Data Forge to demonstrate reusable instructions and independent review. Skills are packaged for different clients from shared source files, while credentials, permissions and executable tests remain separate responsibilities.

Sharing knowledge across clients

GitHub’s November 12 guide uses the public Proof Agent and Data Forge examples to explain reusable engineering instructions centered on SKILL.md. Data Forge discovers project schemas and naming conventions, writes a manifest, and shares that context among building, platform-access and linting skills.

The example packages shared skill sources for Claude Code, Cursor and GitHub Copilot. Task knowledge travels between clients; models, tools, credentials and permissions still depend on the host. Instructions for querying a database do not grant access.

Evidence before a verdict

Proof Agent gives an independent verifier the original request and code diff. Findings must cite evidence in changed files and return PASS, FAIL or PARTIAL. GitHub explicitly says the example Action performs static review without executing code or tests. Its demonstration configuration makes failures advisory rather than merge-blocking.

The practical distinction is between reusable guidance and enforced checks. Teams can share skills while configuring actual merge and release requirements in CI and repository rules; an agent’s completion statement cannot supply those guarantees.

via: GitHub: practical guide to agent skills