GitHub showed teams how to stop re-explaining the same job to every agent at Universe 2026. Andrea Griffiths presented the method live with Shishir Tewari, Senior Manager, Data Engineering at Procore.
A skill is instructions, not access.
An agent skill is a versioned directory centered on a SKILL.md file that tells an agent when and how to do one repeatable task. The file opens with frontmatter that names the skill and describes its trigger, so the agent loads detail only when relevant. For Proof Agent, that frontmatter declares adversarial verification of AI-generated work and a central rule that self-verification is not verification. The body then defines when to verify and what evidence the verdict must cite.
The Data Forge support-desk example shows how small skills combine into a larger run. The forge-init skill discovers source tables and naming conventions and writes them to a resolved manifest. The forge-builder skill then drives the work from issue to pull request, calling forge-dbx for queries and forge-lint for team standards. Each skill keeps one responsibility and reads shared project context instead of rediscovering it.
Project detail stays out of the skill. The support-desk manifest names the project, the support_metric_store and a DuckDB engine with a spark_python pipeline framework, according to the guide. Another team can reuse the same skills with its own manifest, which is the portability test. That reuse stays cheap because agents load only name and description at startup, about 100 tokens per skill, and pull the full body only on activation, per an independent SKILL.md explainer.
Checks sit in the workflow, not in the model. Data Forge runs seven phases with a human checkpoint at each one and approval before schema changes run. Proof Agent gets the original request and the diff, but not the coding agent assessment, and must return PASS, FAIL or PARTIAL with citations in changed files. The demo set block-on-fail to false for advisory findings, while by default a FAIL fails the check and can block merge if required.
Skills fit bounded work like checklists, releases and metric pipelines. When you need a stable interface, shared state, an audit trail or permissions enforced outside the model, build an application and watch whether manifests stay reviewable.
