claude-resources-share

name	claude-resources-share
description	Share Claude Code resources (memories, settings, skills, hooks, etc.) across projects via the centralized claude-settings repo. Use when: (1) User says 'share claude resource', 'sync settings', or 'export to claude-settings', (2) User wants to copy .claude/* files to the central repo, (3) Reusing memory or skills across projects, (4) Backing up local Claude config to the global repo.
argument-hint	[-a|--auto]

Claude Resources Share

One-direction publish from $HOME/.claude/ (private) to $HOME/repos/p/claude-resources (public).

Auto mode

If the user passes -a or --auto (e.g. /claude-resources-share -a, /claude-resources-share --auto, or wording like "auto", "if no problems do copy and push", "do it all if clean"), enable auto mode:

• Step 1: still run the scan, but if findings are clean (no HIGH/MEDIUM priority issues per /purge-private-info), skip the user-confirmation gate and proceed automatically. If findings are NOT clean, stop and report — do not auto-proceed.
• Step 6: skip the choice prompt and run /commits with push (option 1) automatically.

In normal mode (no flag), follow the original gates: ask for confirmation after the scan and ask which commit option to use.

Paths

• Source: $HOME/.claude/
• Target: $HOME/repos/p/claude-resources

Directories to copy

Source	Target
$HOME/.claude/commands/	commands/
$HOME/.claude/skills/	skills/
$HOME/.claude/agents/	agents/
$HOME/.claude/hooks/	hooks/
$HOME/.claude/scripts/	scripts/
$HOME/.claude/web/	web/
$HOME/.claude/CLAUDE.md	CLAUDE.md

Excludes (apply to all rsync operations)

• node_modules/
• .claude/ (stray nested config dirs inside subtrees, e.g. skills/.claude/)
• .docusaurus/
• build/
• dist/
• __pycache__/
• .DS_Store
• *.pyc
• .cache/
• pnpm-lock.yaml
• package-lock.json
• target/

Symlinks

Skip symlinks during copy. Symlinks in $HOME/.claude/skills/ typically point to project-local repos that do not exist on a fresh machine, so dereferencing or preserving them in the public bundle would break the install. The rsync flag --no-links skips them entirely.

Plugin manifests (preserved)

The public repo contains .claude-plugin/marketplace.json and .claude-plugin/plugin.json so users can install the bundle via /plugin marketplace add takazudo/claude-resources. These files live ONLY in the public repo (not in $HOME/.claude/), and Step 3's cleanup loop must NOT delete .claude-plugin/.

Workflow

Step 1: Scan for private info

Invoke the /purge-private-info command, targeting the source directories listed above. Scan file contents for:

• API keys, tokens, passwords, webhook keys
• Client/corporate names that appear to be real clients
• Personal SNS accounts (other than the repo owner's)
• Hardcoded absolute paths containing usernames (e.g., $HOME/)

Present all findings to the user. Do NOT proceed until the user explicitly confirms there are no problems or that findings are acceptable.

In auto mode (-a / --auto): if the scan reports no HIGH/MEDIUM priority findings (clean result), skip the confirmation gate and proceed to Step 2. If anything is flagged, stop and report — auto mode never auto-confirms a non-clean scan.

Step 2: Prepare target directory and pull latest

# Create target if it doesn't exist
mkdir -p $HOME/repos/p/claude-resources

If the target has a .git/ directory, pull latest to avoid conflicts:

cd $HOME/repos/p/claude-resources
git pull --rebase

If not a git repo, the user should initialize it separately.

Step 3: Remove old content and fresh copy

First, remove all old content in the target (preserving .git/, .gitignore, README.md, LICENSE):

# Remove previous copies (but preserve git and repo meta files)
cd $HOME/repos/p/claude-resources
for dir in commands skills agents hooks scripts web; do
  rm -rf "./$dir"
done
rm -f ./CLAUDE.md

Then copy fresh from source using rsync. IMPORTANT: Pass each --exclude flag separately — do NOT store them in a shell variable, as variable expansion breaks glob patterns like *.pyc. The --no-links flag is critical: it skips symlinked skills/commands/agents that point to project-local repos.

SRC="$HOME/.claude"
DST="$HOME/repos/p/claude-resources"

for dir in commands skills agents hooks scripts web; do
  rsync -av --no-links \
    --exclude=node_modules \
    --exclude=.claude \
    --exclude=.docusaurus \
    --exclude=build \
    --exclude=dist \
    --exclude=__pycache__ \
    --exclude=.DS_Store \
    --exclude='*.pyc' \
    --exclude=.cache \
    --exclude=pnpm-lock.yaml \
    --exclude=package-lock.json \
    --exclude=target \
    "$SRC/$dir/" "$DST/$dir/"
done
cp "$SRC/CLAUDE.md" "$DST/CLAUDE.md"

Step 4: Verify and clean

After copying, verify no excluded artifacts leaked through and confirm the plugin manifests are still in place:

DST="$HOME/repos/p/claude-resources"

# Check for leaked node_modules
leaked=$(find "$DST" -name node_modules -type d 2>/dev/null)
if [ -n "$leaked" ]; then
  echo "WARNING: Leaked node_modules found, removing:"
  echo "$leaked"
  find "$DST" -name node_modules -type d -exec rm -rf {} + 2>/dev/null
fi

# Clean .DS_Store files
find "$DST" -name .DS_Store -delete 2>/dev/null

# Sanity check: plugin manifests must exist for marketplace install to work
for f in .claude-plugin/marketplace.json .claude-plugin/plugin.json; do
  if [ ! -f "$DST/$f" ]; then
    echo "ERROR: Missing $f -- marketplace install will fail. Restore before pushing."
  fi
done

# Sanity check: no symlinks should have leaked through
leaked_links=$(find "$DST" -type l 2>/dev/null)
if [ -n "$leaked_links" ]; then
  echo "WARNING: Symlinks leaked through (should be skipped via --no-links):"
  echo "$leaked_links"
  find "$DST" -type l -delete 2>/dev/null
fi

Step 5: Summary

Report file counts per directory:

DST="$HOME/repos/p/claude-resources"
for dir in commands skills agents hooks scripts web; do
  echo "$dir: $(find "$DST/$dir" -type f | wc -l) files"
done
echo "CLAUDE.md: 1 file"

Step 6: Commit and push

Normal mode: Ask the user whether they want to commit and push the changes to the target repo. Do NOT auto-commit or auto-push. Present the options:

1. Commit and push
2. Commit only (no push)
3. Skip (leave changes uncommitted)

If the user chooses to commit, use the /commits skill to commit inside the target repo directory. If they also want to push, push after committing.

Auto mode (-a / --auto): skip the prompt. Run /commits inside the target repo directory and push (equivalent to option 1) without asking. Report the commit and push result back to the user.

Scan whitelist

The following items are known and acceptable — do NOT flag them during the Step 1 scan:

• IFTTT webhook event names in hooks/notify-ifttt.sh (the key itself is in an env var)
• References to "CodeGrid", "esa", repo names like takazudo-codegrid-writing, takazudo-esa-writing in skills/agents (publicly known authorship)
• $HOME/repos/w/ and $HOME/repos/p/ directory structure references (personal convention, no secrets)

Important rules

• Never copy from public to private. This is strictly one-direction.
• Never skip Step 1. The safety scan is mandatory every time.
• Never auto-confirm a non-clean scan. In auto mode, only proceed when the scan is fully clean; any HIGH/MEDIUM finding stops the flow and requires user input. In normal mode, always wait for explicit user approval after the scan.
• Never store rsync excludes in a shell variable. Pass each --exclude flag inline to avoid glob expansion issues.
• Never delete .claude-plugin/ during cleanup. Step 3 only removes commands/, skills/, agents/, hooks/, scripts/, web/, and CLAUDE.md — the plugin manifests are public-repo-only files and must persist across shares.
• Never include symlinks. Use --no-links in rsync. Symlinked skills/commands/agents point to project-local repos that do not exist on a fresh machine and would break installs from the marketplace.