github-copilot-agent-skills
Use for creating, reviewing, debugging, or packaging portable GitHub Copilot Agent Skills with SKILL.md, precise triggers, progressive disclosure, scripts, examples, and resources. Do not use for alwa
SKILL.mdskills/github-copilot-agent-skills/SKILL.md
GitHubGitHub Copilot Agent Skills
Build a small, portable execution contract that Copilot can discover from its metadata and execute without reconstructing expert judgment. Inspect the installed version and active host schema before using optional fields.
Workflow
- Define one repeated job, inputs, artifacts, invariants, completion evidence, positive triggers, boundary cases, and near-miss negatives.
- Choose project scope (
.github/skills,.agents/skills, or compatible location) or explicitly requested personal scope (~/.copilot/skills). - Create a lowercase kebab-case directory and matching
SKILL.mdname. Front-load thedescriptionwith the job and nearest non-goals. - Write decision rules, ordered workflow, canonical examples, and verification. Put optional stable detail in directly linked resources.
- Add deterministic scripts only when they remove repeated error-prone logic; document input, output, failure, dependencies, and safety.
- Test explicit slash use, implicit discovery, baseline improvement, negative interference, and resource access separately.
Rules
- Required metadata is
nameanddescription. Current VS Code also supportsargument-hint,user-invocable,disable-model-invocation, and experimentalcontext: fork; verify portability before using optional fields. - Keep every normal resource one hop from
SKILL.mdand reference it explicitly. - Use manual-only invocation when automatic matching would be unsafe or noisy.
- Never put authoring research, evaluation fixtures, secrets, or mutable status in the distributed skill.
- Prefer Python 3.11+ standard-library helpers when executable logic is needed.
Read the skill contract, use skill patterns, and apply verification.
Completion
The name matches its folder, metadata and links validate, runtime content is minimal and self-contained, helpers are tested, discovery boundaries are measured, and VS Code exposes the skill in the intended project or user scope.