GitHub's practical guide to agent skills: build once, run on any agent

GitHub Blog AI/ML · rss · 2026-11-13

GitHub's official blog walks through turning repeatable engineering judgment into versioned Agent Skills built around a SKILL.md file, reusable across agents and clients.

Key ideas

A skill captures instructions, not access: the environment still provides tools, credentials, and permissions.
Core rule: self-verification is not verification — the Proof Agent example spawns an independent verifier that returns PASS/FAIL/PARTIAL with evidence cited from changed files.

Coordinating multiple skills (Data Forge)

forge-init discovers tables and conventions into a resolved manifest; forge-builder, forge-dbx, and forge-lint all read that manifest instead of re-discovering context.
forge-dbx gives every skill one consistent path to the data platform, selecting a Databricks or DuckDB adapter via the manifest.

Three practices

Put project config in one manifest file rather than hardcoding;
Give related skills a single consistent execution path;
Use instructions to guide work and checks to verify it, with human approval points and CI-enforced merge blocks where needed.

Packaging: keep skills//SKILL.md as the source and generate client-specific files for Claude Code, Cursor, and Copilot via a script with a --check mode for CI. The guide also notes skills fit bounded, repeatable work, while workflows needing stable interfaces, audit trails, or enforced permissions are better as applications.

Original post →

GitHub's practical guide to agent skills: build once, run on any agent

More from coding & agent