Skip to main content

Agent Runtime Skills

Agent runtime skills are the files your AI tooling actually loads. They do not need to contain every rule; they should point the agent to the right organization and project skills for the current task.

For Copilot, these runtime files are the practical entry point. They shape chat responses, inline suggestions, and agentic edits by pointing Copilot at the skills it must follow.

Example files:

project/SKILLS.md
project/.github/copilot-instructions.md
project/.cursor/rules/frontend.md
project/.claude/CLAUDE.md

Short content:

# project/SKILLS.md

Load organization skills from `.skills/engineering-skills`.
Load project skills from `.skills/project`, `.skills/domains`, and `.skills/features`.
Ask for missing domain rules before generating code.

Think of the runtime file as the active skill set. It selects the skills the agent should follow, combines shared standards with project context, and keeps the prompt surface small.

Example files:

# SKILLS.md

Required Skills:

- .skills/engineering-skills/react/page.md
- .skills/engineering-skills/quality/testing.md
- .skills/project/architecture.md
- .skills/domains/customer-management.md

Short content:

# .skills/project/architecture.md

Use feature folders for product capabilities.
Keep domain logic framework-independent.
Validate generated code with tests before opening a pull request.

Different tools load different files, but the responsibility is the same: give the agent a short, explicit entry point into the skills it must obey.

This keeps AI-driven workflows predictable because the agent starts from a known rule set before it generates or changes code.

Example files:

.github/copilot-instructions.md
.cursor/rules/skills.md
.claude/skills/frontend.md

Short content:

# .cursor/rules/skills.md

Read `.skills/SKILLS.md` first.
Prefer existing components and utilities.
Do not introduce new architectural patterns without updating the relevant skill.

Mini exercise​

Apply Agent Runtime Skills to a feature or workflow you know. In ten minutes:

  1. Describe the current approach in one sentence.
  2. Identify one ambiguity, coupling risk, or missing constraint.
  3. Write one concrete rule or artifact that would improve predictability.
  4. Define how a reviewer or automated check could validate the improvement.

Compare your result with a teammate and revise the rule if two people could interpret it differently.

Example​

For Agent Runtime Skills, consider an accessibility-testing skill:

  • Current approach: The same guidance exists in user prompts, repository notes, and local agent configuration.
  • Risk: Teams cannot tell which version is authoritative.
  • Concrete rule: Store reusable accessibility policy at organization level, project exceptions in the repository, and runtime invocation details with the agent.
  • Validation: Trace one generated component to the exact policy, project rule, and runtime instruction it used.