Project Initialization

Improve the results of running the /init command to create AI agent rules.

When to Use This Skill

Use this skill when:

• Setting up a new repository for AI agent collaboration
• Neither CLAUDE.md nor AGENTS.md exists
• The /init command has been run but needs improvement
• You want AI agent rules to be discoverable by multiple agents

Goal

The /init command creates CLAUDE.md, but this is only useful for Claude Code. This skill ensures the same file contents are discoverable and readable by multiple AI agents looking in different places.

This follows the standard documented at: https://agent-rules.org/

Process

Perform the following steps:

1. Check existing files: Verify if CLAUDE.md and/or AGENTS.md exist

2. Handle new project: If neither file exists, first run Claude's /init process to generate initial content

3. Rename if needed: If only CLAUDE.md exists, rename it to AGENTS.md

4. Create symlinks: Generate symlinks so different agents can find the rules:

   • CLAUDE.md → points to AGENTS.md

5. Fix unsafe git push commands: Scan all agent instruction files (AGENTS.md, CLAUDE.md, and any other files they reference) for bare or underspecified git push invocations that could accidentally push to the wrong branch. This is a common problem with instructions added by tools like beads.

   What to look for: any git push that does NOT specify both a remote and a refspec, e.g.:

   • git push (no arguments at all)
   • git push origin (remote but no refspec)
   • git push --force (flags but no remote/refspec)

   Why this matters: a bare git push relies on the upstream tracking branch configuration. If you're on a feature branch that happens to track main, git push will push your feature commits directly to main. Using git push origin HEAD is much safer because it pushes the current branch to a remote branch of the same name, preventing accidental pushes to trunk branches.

   How to fix: replace bare git push with git push origin HEAD. Preserve any flags that were present, e.g. git push --force → git push origin HEAD --force.

   Edge cases to handle:

   • Lines inside code blocks (```) and inline code (`)
   • Lines in checklists like [ ] 6. git push (push to remote)
   • Comments after the command, e.g. git push # deploy
   • Multiple occurrences in the same file
   • Do NOT modify lines that already specify both remote and refspec (e.g. git push origin main, git push origin HEAD)

6. Verify: Confirm the symlinks work by checking file accessibility

Context

• Existing files: ls CLAUDE.md AGENTS.md

Example Commands

# Check existing files
ls -la CLAUDE.md AGENTS.md

# Rename if needed
mv CLAUDE.md AGENTS.md

# Create symlinks
ln -s AGENTS.md CLAUDE.md

# Verify
cat CLAUDE.md

Fixing unsafe git push

Search agent instruction files for bare git push:

# Find bare git push commands in agent files
grep -n 'git push' AGENTS.md CLAUDE.md .beads/context/*.md

Example transformations:

Before	After
git push	git push origin HEAD
git push --force	git push origin HEAD --force
git push origin	git push origin HEAD
[ ] 6. git push	[ ] 6. git push origin HEAD

Leave these unchanged (already safe):

Already safe
git push origin HEAD
git push origin main
git push origin HEAD --force
git push -u origin HEAD

Result

After this process:

• AGENTS.md contains the authoritative AI agent rules
• CLAUDE.md is a symlink to AGENTS.md
• Multiple AI agents can discover and read the same rules
• All git push commands in agent instructions explicitly specify remote and refspec to prevent accidental pushes to trunk branches