Skill Official Updater

Official Anthropic Skills Updater

Keeps a local copy of Anthropic's official skills (https://github.com/anthropics/skills, under skills/) in sync with this repo's skills_library/, without silently overwriting unrelated or hand-edited skills.

State file

skills_library/.official-skills-state.json (create if missing):

{
  "repoSha": "<last-synced commit SHA of anthropics/skills>",
  "skills": {
    "pdf": "<commit SHA of skills/pdf/ at last sync>",
    "docx": "<commit SHA of skills/docx/ at last sync>"
  }
}

Only skill names listed under skills are "managed" by this updater - a same-named skill in skills_library/ that ISN'T listed there is assumed to be a hand-written/customized skill and must never be overwritten without an explicit user decision.

1. Cheap check (every session)

Automation: When the Claude Skills extension is active and this workspace has skills_library/, it installs a Claude Code SessionStart hook (official-skills-watch.js) that runs the check below on startup, resume, and clear. If updates exist, the hook injects context telling you to read this skill and sync immediately — do not ask the user which skills to pull.

Manual / fallback (same logic):

git ls-remote https://github.com/anthropics/skills.git HEAD

Compare the returned SHA to repoSha in the state file (treat a missing file as "no state yet"). If unchanged, say nothing further - this step should be silent and near-instant.

VS Code: Claude Skills: Check Official Anthropic Skill Updates runs the same check from the Command Palette. Enable Official Skills Session Check re-installs the SessionStart hook if needed.

2. When the repo SHA has changed (or no state file yet)

List the current skill directories upstream:

gh api repos/anthropics/skills/contents/skills --jq '.[].name'

For each upstream skill name, compare against skills_library/:

• New - upstream name not present locally: candidate to add.
• Managed + changed - name is in the state file's skills map AND git log -1 --format=%H -- skills/<name> (in a shallow clone, see below) differs from the stored SHA: candidate to update.
• Unmanaged - a local skill with the same name exists but isn't in the state file's skills map: skip, and report it as "name collision, not touched" so the user is aware but nothing changes automatically.

Auto-sync policy (default — do not ask the user):

• Pull every new and managed + changed candidate automatically.
• Never overwrite an unmanaged collision (same local name, not in state skills map) — list those in the completion summary only.
• Do not stop to ask "which skills" or "all 17?" — proceed with the sync unless git/gh is unavailable or the sparse clone fails.

Briefly note what you are syncing (counts + names), then run the pull.

3. Pulling skills (all actionable candidates)

Use a temporary shallow sparse clone so only the needed skill directories are fetched. On Windows, do not pass -q to git sparse-checkout set.

git clone --depth 1 --filter=blob:none --sparse https://github.com/anthropics/skills.git <tmpdir>
cd <tmpdir>
git sparse-checkout set skills/<name1> skills/<name2> ...

For each actionable skill (new or managed + changed), copy <tmpdir>/skills/<name>/ to skills_library/<name>/. Delete <tmpdir> afterwards.

To record per-skill SHAs for managed skills, after checkout run git log -1 --format=%H -- skills/<name> inside <tmpdir> for each synced skill.

4. Manifest entries

For each newly added skill, add an entry to skills_library/manifest.json:

• description: the skill's SKILL.md frontmatter description, trimmed to one line (unwrap YAML |- blocks to a single line).
• detect_globs: use the table below; do not ask the user.

For updated skills, refresh the manifest description if the upstream frontmatter changed meaningfully.

5. Update state and finish

• Write the new repoSha and per-skill SHAs (for every skill added or updated) to skills_library/.official-skills-state.json.
• If extension/ exists, run npm run sync-skills inside extension/ to refresh the bundled copy (or tell the user if you cannot run commands).
• Summarize what was synced: added count, updated count, skipped collisions.
• If this workspace uses the deployer extension, note that Install Skill Library (or py generate_skills.py install) republishes to ~/.claude/skills/ — run or suggest only if skills were actually changed.

6. No git/gh available

If neither git nor gh is available, say so and stop - don't attempt to fetch individual files via raw URLs as a substitute, since that won't reliably capture a skill's full directory (scripts, references, assets).