Cross Agent Skill Installer

The Problem

Every coding agent stores skills in a different directory. Claude Code reads ~/.claude/skills/, Cursor reads .cursor/skills/, Codex CLI reads ~/.codex/skills/, Copilot reads ~/.copilot/skills/, and OpenClaw reads ~/.claw/skills/. When you install a skill by hand into all five, you get five copies that can drift apart after updates, break relative references to scripts/ or references/, and waste disk space. The new agents/ vendor-neutral convention was supposed to fix this, but nobody agrees on the path yet. This skill gives your agent a repeatable, idempotent install-and-verify procedure that works across all five harnesses today.

What You Get

• Harness discovery map — the exact skill directory for every major agent (Claude Code, Cursor, Codex CLI, Copilot, OpenClaw, and the emerging .agents/ vendor-neutral standard), verified against current documentation and filesystem probing.
• Symlink/copy installer pattern — a reproducible install procedure that stages skills from a canonical source repo, creates symlinks (or copies for agents that don't support symlinks), and records every install in a manifest for idempotent updates.
• Drift detection — a verification step that compares installed skill checksums against the canonical source, so you know immediately when a local copy has diverged or a relative reference has broken.
• Manifest-driven uninstall — clean removal of any skill from any harness by reading the install manifest, with no orphaned directories or stale references.
• Multi-machine sync workflow — a procedure for keeping your skill library consistent across your laptop, desktop, and CI environment.

Why This Beats Prompting It Yourself

Ask an LLM "where does Cursor store skills?" and you get a plausible-sounding path that might be from six months ago. Paths change between versions, some agents accept multiple locations, and the vendor-neutral .agents/ convention is still being finalized. This skill probes the actual filesystem, reads current documentation, and maintains an updated canonical map. The install procedure is mechanical and verified — no guesswork, no "it should work."

Example Session

You: "Install the react-hooks-auditor skill into Claude Code, Cursor, and Codex CLI on this machine."

Agent:

CROSS-AGENT SKILL INSTALL — react-hooks-auditor

Step 1: Discovery
  ~/.claude/skills/    → exists ✓
  .cursor/skills/      → exists ✓
  ~/.codex/skills/     → exists ✓

Step 2: Staging
  Canonical source: ~/skills/react-hooks-auditor/
  Checksum: sha256:a3f8c2...
  Staged to: /tmp/skill-staging/react-hooks-auditor/

Step 3: Install
  ~/.claude/skills/react-hooks-auditor → symlink ✓ (targets canonical)
  .cursor/skills/react-hooks-auditor   → symlink ✓ (targets canonical)
  ~/.codex/skills/react-hooks-auditor  → copy ✓ (symlinks not supported)

Step 4: Verify
  All 3 installs: checksums match ✓, SKILL.md readable ✓, scripts/ accessible ✓
  Relative refs (references/*, scripts/*): all resolve ✓

Manifest written: ~/.agents/install-manifest.json
  3 installs recorded, 0 orphans detected

Use Cases

• Setting up a fresh dev machine with your standard skill library across multiple agents.
• Sharing a custom skill with teammates who use different coding agents.
• Updating a skill in one canonical location and propagating changes to all harnesses.
• Auditing an existing multi-agent setup for stale copies or broken references.
• Migrating from a single-agent workflow to a multi-agent setup without losing skills.

Known Limitations

This skill installs and syncs skills; it doesn't create or evaluate skill quality. Some agents may require a restart before newly installed skills are recognized. The vendor-neutral .agents/ standard is still evolving — paths may change. Symlinks don't work on Windows; the skill falls back to copies with a note that manual sync is needed after updates.