AI guidance governance
June 11, 2026 · View on GitHub
Purpose
This repo targets GitHub Copilot and Claude Code with a minimal, two-tool documentation model:
- Shared standing guidance lives in root and nested
AGENTS.mdfiles. - GitHub-required custom instructions stay in
.github/only where GitHub requires exact paths. - Claude-only workflows live in
.claude/skills/. - Agent definitions in
.github/agents/and role chatmodes in.github/chatmodes/describe behavior/persona, not system architecture.
Source of truth
- Shared repo workflow:
AGENTS.mdplus the nearest nestedAGENTS.md - Component architecture & entry points:
Src/<Component>/AGENTS.md - GitHub Copilot custom instructions:
.github/copilot-instructions.mdand.github/instructions/*.instructions.md - Claude Code workflows:
.claude/skills/*/SKILL.md
No duplication rule
- Do not duplicate shared standing guidance across
AGENTS.md,CLAUDE.md, and GitHub compatibility files. - Keep each Claude skill in exactly one place:
.claude/skills/. - GitHub prompt files may reference Claude skill paths, but do not keep a
mirrored
.github/skills/tree. - If a rule must be auto-applied by GitHub Copilot, keep it in
.github/instructions/*.instructions.md; otherwise prefer the relevantAGENTS.mdor a Claude skill.
What goes where
AGENTS.md
Use for:
- One-page onboarding for both tools: build/test commands, repo constraints, and links.
- Stable, always-on guidance that should apply across sessions.
.github/instructions/*.instructions.md
Use for:
- Prescriptive constraints that must be applied during editing/review.
- Cross-cutting rules that GitHub Copilot needs through its official, path-scoped instruction mechanism.
Curated keep set (intentionally small):
build.instructions.mddebugging.instructions.mdinstaller.instructions.mdmanaged.instructions.mdnavigation.instructions.mdnative.instructions.mdpowershell.instructions.mdrepo.instructions.mdsecurity.instructions.mdterminal.instructions.mdtesting.instructions.md
Src/**/AGENTS.md
Use for:
- Where to start (entry points, key projects, typical workflows).
- Dependencies and cross-component links.
- Tests (where they live, how to run them).
Baseline expectations for a component agent doc:
- Where to start (projects, primary entry points)
- Dependencies (other components/layers)
- Tests (test projects and the recommended
./test.ps1invocation)
.claude/skills/
Use for:
- Claude-only workflows, repeatable procedures, and task-specific guidance.
- Put multi-step procedures here instead of inflating
CLAUDE.md.
.github/agents/ and .github/chatmodes/
Use for:
- GitHub Copilot role definitions, boundaries, and tool preferences.
- Do not put component architecture here; link to the component
AGENTS.md.
External standards alignment (post-2025)
- AGENTS.md: Supported as a simple, vendor-neutral instruction format by multiple tools (for example, Cursor’s project rules). Use plain Markdown with clear headings and concise rules.
- MCP (Model Context Protocol): Use MCP for tool/data integration rather than vendor-specific plugins; MCP provides a standardized, versioned protocol for AI tool connectivity.
Adding a new scoped instruction file
Add a new .github/instructions/<name>.instructions.md only when:
- The guidance is prescriptive (MUST/DO NOT), and
- It applies broadly or to a subtree, and
- It would be harmful if Copilot ignored it.
Otherwise, update the appropriate AGENTS.md or add a Claude skill under
.claude/skills/.