The development

What happened

GitHub's 12 November practical guide explains how AI coding-agent skills can share reusable instructions while tools, project configuration and permissions remain specific to each host. Its Data Forge and Proof Agent examples distinguish portable files from a working, enforced workflow.

Key facts

  • A skill is a versioned directory centered on SKILL.md. In the Data Forge example, a resolved manifest supplies project details, while a separate linting skill provides team standards; discovered repository patterns are not automatically rules to preserve.

  • Data Forge's example workflow has seven phases with human checkpoints and asks for approval before schema changes. A manifest can select a database adapter, but cannot install its tools or grant database access.

  • Proof Agent's GitHub Action gives a separate verifier the original request and diff, excluding the coding agent's own assessment. It performs static review, not code execution or tests. The demonstration makes findings advisory; a failed check blocks merging only when repository rules require it.

  • Data Forge packages shared skill files for Claude Code, Cursor and GitHub Copilot. GitHub advises checking that each client recognizes the skill, loads its references, can use the required tools and returns the requested evidence.

The context

Why it matters

Reusable instructions can reduce repeated explanations without making agent environments interchangeable. Teams adopting them still need to verify client setup and enforce permissions, tests and merge requirements outside the prose. This is GitHub's practical guidance, not a new product launch or an independently tested claim of universal portability or secure code.