Github Copilot Instructions

GitHub Copilot instructions

Encode stable repository facts and conventions at the narrowest automatic scope. Instructions affect chat requests, not inline suggestions.

Workflow

1. Inspect existing instruction files, project configuration, build/test entry points, nested repositories, and the active VS Code settings.
2. Separate rules that apply to nearly every task from language, framework, or directory-specific rules.
3. Put broad rules in .github/copilot-instructions.md. Put conditional rules in .github/instructions/<purpose>.instructions.md with applyTo globs.
4. Use AGENTS.md only for durable cross-agent repository guidance. Treat nested discovery as host/setting dependent.
5. Rewrite every statement as fact, condition/action, prohibition, or exact verification command. Remove aspiration, task-specific prompts, and history.
6. Check overlaps. Applicable files can be combined without guaranteed order; eliminate contradictions rather than declaring a winner.

Rules

• Keep always-on context short, self-contained, and broadly applicable.
• Ground commands in files that exist and commands that were actually tested.
• Use forward-slash repository-relative applyTo patterns. Test positive and negative paths; do not assume *.py recursively matches the repository.
• Link local context with relative Markdown links rather than copying it.
• Do not put secrets, mutable status, long API references, personas, or one-off workflow prompts in instructions.
• Do not use deprecated settings-based code/test-generation instructions for a new design. Inspect current host settings for review/commit/PR special cases.

Read scope and glob contracts. Use instruction patterns and finish with verification.

Completion

Every rule has a justified scope; YAML parses; globs select intended files and exclude near misses; commands and paths exist; no instruction sets conflict; and VS Code lists the applied instruction reference in a representative chat.