| name | ai-instruction-standards |
| scope | partial |
| description | Create and maintain AI instruction files (CLAUDE.md, AGENTS.md, .cursor/rules/, etc.) with proper structure.
Use when: creating AI instruction files, separating universal vs project-specific rules, configuring AI tools.
Keywords: CLAUDE.md, AGENTS.md, cursorrules, windsurfrules, clinerules, AI instructions, system prompt.
|
AI Instruction File Standards Guide
Language: English | 繁體中文
Version: 2.0.0
Last Updated: 2026-04-28
Applicability: All AI coding tools
Core Standard: This skill implements AI Instruction File Standards. For comprehensive methodology documentation, refer to the core standard.
AI Skills Hierarchy | AI 技能層級
This skill is part of a three-layer AI collaboration system:
| Layer | Skill | Question it Answers |
|---|
| Behavior (Immediate) | /ai-collaboration | "How should AI respond accurately?" |
| Configuration (Session) | /ai-instruction-standards (this) | "What to write in CLAUDE.md / AGENTS.md?" |
| Architecture (Long-term) | /ai-friendly-architecture | "How to structure code for AI?" |
Purpose
This skill helps create and maintain AI instruction files with proper separation between universal standards and project-specific configurations, across all major AI coding tools.
Quick Reference
Supported AI Tools (2026-04-28)
CLI / Agent Tools (Terminal)
| Tool | Primary File | Workflow Mechanism | MCP |
|---|
| Claude Code | CLAUDE.md + .claude/rules/*.md | Skills (.claude/skills/ → /{name}) | ✅ |
| Gemini CLI | GEMINI.md | .gemini/commands/*.toml → /{name} | ✅ |
| OpenAI Codex CLI | AGENTS.md (+ AGENTS.override.md) | Team commands; /review built-in | ✅ |
| OpenCode | AGENTS.md (CLAUDE.md compatible) | Built-in only (/init /undo /share) | ✅ |
AI-native IDE / Editor Integration
| Tool | Primary File | Workflow Mechanism | MCP |
|---|
| Cursor | .cursor/rules/*.mdc ⚠️ | @-mentions; /multitask | ✅ |
| GitHub Copilot | .github/copilot-instructions.md | .github/prompts/*.prompt.md → /{name} | ✅ |
| Windsurf | .windsurfrules / .windsurf/rules/*.md | .windsurf/workflows/*.md → /{name} | ✅ |
| Cline | .clinerules | None | ✅ |
⚠️ Cursor: .cursorrules is deprecated — migrate to .cursor/rules/*.mdc
Cross-Tool Universal Standard: AGENTS.md
AGENTS.md is the emerging de-facto cross-tool instruction standard:
Supported by: Gemini CLI, OpenAI Codex CLI, OpenCode, GitHub Copilot, Windsurf, Cursor
Not supported by: Claude Code (uses CLAUDE.md), Cline (uses .clinerules)
Recommendation: Use AGENTS.md as the universal baseline for cross-tool projects, then add tool-specific files for advanced features (Skills, Workflows, Prompts).
Core Principle: Universal vs Project-Specific
| Type | Contains | Example |
|---|
| Universal | Generic rules | "Run tests before committing" |
| Project-Specific | Concrete commands | "Run npm test before committing" |
Recommended Layout
# [Project Name] - AI Instructions
## Universal Standards
<!-- Rules applicable to ANY project -->
- Commit message format
- Code review checklist
- Testing standards
- Anti-hallucination rules
---
## Project-Specific Configuration
<!-- Unique to THIS project -->
### Tech Stack
[Your technologies here]
### Quick Commands
[Your build/test/deploy commands]
### File Structure
[Your project structure]
Tool-Specific Setup Guides
Claude Code
CLAUDE.md # Main instructions (hierarchical: global → project → subdir)
.claude/rules/ # Glob-scoped additional rules
.claude/skills/{name}/SKILL.md # Custom slash commands → /{name}
.claude/agents/{name}.md # Subagent definitions
Gemini CLI
GEMINI.md # Main instructions
.gemini/commands/{name}.toml # Custom slash commands → /{name}
.gemini/agents/{name}.yaml # Subagent definitions
Example .gemini/commands/review.toml:
description = "Run code review checklist"
prompt = "Review the following changes: !{git diff HEAD}"
OpenAI Codex CLI
AGENTS.md # Main instructions (Git root → cwd traversal)
AGENTS.override.md # Temporary override (highest priority)
~/.codex/AGENTS.md # Global fallback
.codex/agents/ # Custom agent definitions
OpenCode
AGENTS.md # Primary (auto-recognized)
CLAUDE.md # Also recognized (migration compatibility)
.opencode/agents/ # Custom agent definitions
opencode.json (instructions) # Glob-pattern file references
Cursor
.cursor/rules/ # MDC format rules (replaces .cursorrules)
{name}.mdc # Frontmatter: description, globs, alwaysApply
AGENTS.md # Also supported for agent context
MDC frontmatter example:
---
description: "TypeScript coding standards"
globs: ["**/*.ts", "**/*.tsx"]
alwaysApply: false
---
Migration: If you have .cursorrules, move content to .cursor/rules/*.mdc.
GitHub Copilot
.github/copilot-instructions.md # Always-on, all chats
.github/instructions/*.instructions.md # File-glob scoped (applyTo frontmatter)
.github/prompts/*.prompt.md # Reusable templates → /{name} slash commands
.github/agents/*.agent.md # Custom agents with tool access control
AGENTS.md # Also recognized
Windsurf
.windsurfrules # Project rules (team-shareable)
.windsurf/rules/*.md # MDC frontmatter structured rules
.windsurf/workflows/*.md # Reusable task sequences → /{name}
AGENTS.md # Also recognized
Workflow example (.windsurf/workflows/review.md):
Run a code review:
1. Run `git diff HEAD`
2. Check for BLOCKING issues (security, correctness)
3. Check for IMPORTANT issues (design, tests)
4. Output findings with BLOCKING/IMPORTANT/SUGGESTION prefixes
Multi-Tool Project Configuration
When a project uses multiple AI tools:
project/
├── AGENTS.md # Universal baseline (cross-tool)
├── CLAUDE.md # Claude Code (extends AGENTS.md)
├── GEMINI.md # Gemini CLI
├── .cursor/rules/
│ └── standards.mdc # Cursor
├── .windsurf/
│ └── workflows/ # Windsurf workflows
│ ├── review.md
│ └── checkin.md
└── .github/
├── copilot-instructions.md # Copilot always-on
└── prompts/
└── review.prompt.md # Copilot slash command
Best Practice: Write universal content in AGENTS.md once, then import/reference it from tool-specific files to avoid duplication.
Content Guidelines
Universal Content (Keep Generic)
| Category | Good Examples |
|---|
| Commit Standards | "Follow Conventional Commits format" |
| Code Review | "Use BLOCKING, IMPORTANT, SUGGESTION prefixes" |
| Testing | "Maintain 80% coverage minimum" |
| AI Behavior | "Always read code before analyzing" |
Avoid in Universal Sections:
- Specific commands (
npm test, pytest)
- Hardcoded paths (
cli/src/, /var/www/)
- Version numbers (
Node.js 18, Python 3.11)
- Project names and URLs
Project-Specific Content
| Category | Examples |
|---|
| Tech Stack | Node.js 18, React 18, PostgreSQL 15 |
| Commands | npm run lint, ./scripts/deploy.sh |
| File Structure | src/, cli/, tests/ |
| Team Conventions | Traditional Chinese comments |
Maintenance Checklist
Before committing changes to AI instruction files:
Configuration Detection
Detection Order
- Check for existing
CLAUDE.md, AGENTS.md, GEMINI.md, or equivalent files
- Detect which AI tools are in use (check for
.cursor/, .windsurf/, .github/copilot-instructions.md, etc.)
- Analyze content structure for universal/project-specific separation
- If not found, suggest creating structured AI instruction file
First-Time Setup
If no AI instruction file found:
- Ask: "This project doesn't have an AI instruction file. Which AI tools do you use?"
- Recommend
AGENTS.md for cross-tool projects, CLAUDE.md for Claude Code only
- Determine project type and tech stack
- Generate template with appropriate sections
- Add to
.gitignore if contains sensitive info
Next Steps Guidance
After /ai-instruction-standards completes, suggest:
- Create or update project's
CLAUDE.md / AGENTS.md ⭐ Recommended — Apply standards immediately
- Run
/ai-friendly-architecture to optimize AI collaboration at the architecture level
- Run
/ai-collaboration to review AI behavior guidelines
Related Standards
Version History
| Version | Date | Changes |
|---|
| 2.0.0 | 2026-04-28 | Add Gemini CLI, OpenAI Codex CLI; update Cursor (MDC format, deprecated .cursorrules); update OpenCode (AGENTS.md primary); update Copilot (multiple file types); update Windsurf (Workflows); add AGENTS.md cross-tool standard section |
| 1.0.0 | 2026-01-25 | Initial release |
License
This skill is released under CC BY 4.0.
Source: universal-dev-standards