agent-guardrails-template copilot-instructions.md

GitHub Copilot Instructions

These instructions apply to all Copilot completions, suggestions, and chat interactions in this repository.

The Four Laws of Agent Safety

1. Read Before Editing - Never suggest modifications without reading the file first
2. Stay in Scope - Only work on files within the authorized task scope
3. Verify Before Committing - Ensure suggested code compiles, passes lint, and is tested
4. Halt When Uncertain - Ask for clarification instead of guessing

Code Generation Rules

Scope (MANDATORY)

• Only modify files explicitly requested by the user
• Do not refactor unrelated code "while I'm here"
• Do not add new files unless asked
• Do not delete files unless asked
• When scope is unclear: ask for confirmation

Production-First

• Production code must be written before test code
• Tests written against existing production code, not stubs
• Infrastructure code comes after both production and tests

Error Handling

• Validate inputs and handle errors explicitly
• Return meaningful error messages
• Never silently swallow exceptions
• Prefer explicit error returns over panics/exceptions

Security

• Never suggest committing secrets, keys, or credentials
• Never suggest .env files in version control
• Validate all external inputs
• Use parameterized queries, never string concatenation for SQL

Forbidden Patterns

NEVER suggest code that:

• Modifies files that haven't been read
• Mixes test and production environments
• Commits secrets or credentials
• Runs untested code in production
• Makes assumptions about user intent

Three Strikes Rule

When a suggestion is rejected or produces errors:

• 1st failure: Adjust approach and retry
• 2nd failure: Try completely different approach
• 3rd failure: HALT and ask user for guidance

File Headers

Include at the top of new files:

// File: <filename>
// Purpose: <one-line description>
// Created: <date>
// Author: AI-assisted (see git history)

Or equivalent in the target language.