Organize Skills

Organize Skills

Reorganize a skill library so names encode what each skill actually touches, and keep symlinks and cross-references consistent when you rename.

Many setups keep ONE canonical skill library in a shared dir and symlink it into each tool's hub (e.g. ~/.claude/skills/, ~/.codex/skills/). Some skills live inside a project repo and are symlinked into the hub. A single rename can therefore touch several symlinks plus every doc that references the old name — that fan-out is what this skill manages.

Step 1 — Inventory

List every skill and where it really lives. A hub entry is usually a symlink to the canonical dir:

for hub in <hub-1> <hub-2>; do
  echo "== $hub =="; ls -la "$hub"
done

Resolve each name to its real directory with readlink -f. Note which skills live in a project repo versus the shared library — repo-resident renames must be git mv in that repo.

Step 2 — Classify by backend (the naming rule)

The convention: the prefix names the backend the skill depends on, decided by what it actually calls, not by its topic. Grep each skill dir for the tell-tale dependency:

grep -rlE "<api-host>/api|/api/<x>|<db-client>|<AUTH_TOKEN_NAME>" "$skill_dir"

• Calls a specific product's API / database → that product's prefix (e.g. acme-*).
• Pure-local (browser automation, offline SOP, file processing, no backend) → a generic/local prefix.
• Keep a topical infix when a family is large (e.g. acme-seo-* vs acme-reddit-*) so related skills sort together.

State the classification back to the user before renaming. A skill that feels product-specific but makes zero backend calls belongs in the local bucket, and vice versa — go by the dependency.

Step 3 — Rename (and fix everything a rename breaks)

For each rename OLD → NEW:

1. Move the canonical dir.
   • Shared-library-resident: mv <lib>/OLD <lib>/NEW
   • Repo-resident: cd <repo> && git mv <skills-path>/OLD <skills-path>/NEW && git commit
   • Misplaced in the wrong repo? Move it to the shared library rather than just renaming in place.
2. Rewire symlinks in every hub. Remove the stale link, create the new one pointing at the canonical dir. Do this for each tool hub that mirrors the library.
3. Update the name: frontmatter inside NEW/SKILL.md to match the directory — the loader keys on it, not on the folder name.
4. Fix cross-references. Other skills and any long-term memory/notes files may mention OLD by name or by old path:

   grep -rln "OLD" <lib> <memory-dirs> 2>/dev/null
   

   Replace the most-specific string first (full old path before the bare name) so a skill that also moved gets its path updated cleanly.

Step 4 — Verify

# broken symlinks in any hub
for hub in <hub-1> <hub-2>; do
  find -L "$hub" -maxdepth 1 -type l ! -exec test -e {} \; -print
done
# no lingering old names
grep -rn "OLD" <lib> 2>/dev/null
# each renamed skill's name: matches its dir
grep -m1 "^name:" <lib>/NEW/SKILL.md

Confirm the loader re-registers each renamed skill under its new name. Remove any pre-existing dead symlinks you find while sweeping — cheap to fix, confusing to leave.

Notes

• Never drop a skill you didn't rename. Only touch what the convention requires.
• Renames are cheap and reversible; broken cross-references are the real cost — Steps 3.4 and 4 are the point of this skill, not the mv.
• Commit repo-resident renames one logical rename per commit so they bisect and revert independently.