Independent tech intelligence, checked against primary sources.

TechPulseMind Useful technology.
No manufactured hype.

GitHub Guide: Agent Skills That Run on Claude Code, Cursor and Copilot

GitHub's guide shows how agent skills in SKILL.md files can be packaged for Claude Code, Cursor and Copilot, with project details kept in a manifest.

GitHub Guide: Agent Skills That Run on Claude Code, Cursor and Copilot

GitHub has published a practical guide to agent skills — versioned directories built around a SKILL.md file that describe when and how to perform a repeatable task — and shown how one set of instructions can be packaged for Claude Code, Cursor and GitHub Copilot. The post, by Senior Developer Advocate Andrea Griffiths, appeared on November 12, 2026, and follows a GitHub Universe 2026 session she ran with Shishir Tewari, Senior Manager of Data Engineering at Procore.

The core idea is that skills carry instructions, not access. According to GitHub, a skill can describe how to query a database, but it doesn’t grant access to that database — tools, credentials and permissions stay with the host that runs the skill.

One job per skill, and a manifest for project details

GitHub recommends giving each skill a clear job, input and output. Its Proof Agent example uses a separate verifier to check AI-generated changes, on the principle that “self-verification is not verification.” The skill’s frontmatter names the task and describes its purpose.

For larger workflows, the guide’s Data Forge example splits a data pipeline across cooperating skills: forge-init discovers source tables and naming conventions and writes them to a resolved manifest, forge-builder coordinates the work from issue to pull request, forge-dbx queries the data platform, and forge-lint applies team standards. Project names, engines and conventions live in the manifest rather than inside the skills, so another team could reuse them with its own configuration.

GitHub also draws a line around where skills fit. Bounded, repeatable work such as a checklist, release, migration or metric pipeline suits a skill; workflows that need a stable interface, shared state, an audit trail, long-running control or permissions enforced outside the model are better served by an application.

Packaging for each client

To avoid keeping several copies of the same instructions in sync, the Data Forge example keeps skills/*/SKILL.md as the source and uses a packaging script to generate .claude-plugin/plugin.json for Claude Code, .cursor-plugin/plugin.json for Cursor and .github/agents/*.agent.md for GitHub Copilot. The script includes a CI check mode: node scripts/package.mjs --check.

Because clients differ in models, tools, credentials and permissions, GitHub advises checking after installation that the client recognizes the skill for the intended task, loads its references, can use the required tools and returns the evidence the skill calls for.

The guide also puts human approval points inside workflows — the Data Forge example has seven phases, each with a checkpoint — and keeps blocking requirements in the repository. In the Proof Agent demo, the GitHub Action performs a static review of the diff without executing code or running tests, and block-on-fail was set to false so findings were advisory; by default a FAIL fails the check.

The full guide, including the Proof Agent and Data Forge examples, is available on the GitHub Blog, which also links to instructions for adding an agent skill to GitHub Copilot.

Related reading

Sources

Some links on this page may be affiliate links. If you buy through them we may earn a commission at no extra cost to you. See our affiliate disclosure.