skill

Skill Manager

End-to-end workflow for improving Cowork plugin skills: gather feedback, patch code, test, rebuild, and deploy.

Canonical Skill Location

The user's custom skills live in the AGENTS.md git repo:

mnt/Alex/AGENTS.md/skills/

Each skill is a subdirectory (e.g. organize-arxiv/) containing at minimum a SKILL.md and any scripts it references. When deploying patched scripts, always copy them to this location in addition to packaging the .plugin file. This is the source of truth — not OneDrive, not Downloads.

Locating Plugins

Installed plugins live under mnt/.remote-plugins/ (read-only in sandbox) and mnt/.local-plugins/. Plugin source may also exist at the canonical skill path above or a path recorded in auto-memory.

To find a plugin by name:

find mnt/.remote-plugins mnt/.local-plugins -name "plugin.json" -exec grep -l '"name"' {} \; 2>/dev/null

Read plugin.json to confirm identity, then map the full file tree:

find /path/to/plugin -type f

Workflow

1. Gather Feedback

Collect issues from any combination of:

• User input in the current conversation ("the script misidentifies X", "it crashes on Y")
• Auto-memory — read mnt/.auto-memory/MEMORY.md and any project/feedback files for the plugin
• Execution logs — if the user ran the skill recently, check the conversation or ask for error output

Produce a numbered list of issues to fix. Confirm with user before proceeding.

2. Prepare Writable Copy

Plugin dirs under mnt/.remote-plugins/ are read-only. Always:

1. Copy the entire plugin tree to the working directory:

   cp -r /path/to/plugin /sessions/SESSIONID/plugin-name
   

2. All edits happen on this copy.

3. Patch

For each issue:

1. Read the relevant source file(s).
2. Apply the minimal fix. Preserve existing style and formatting.
3. If the fix involves a Python script, prefer defensive changes (try/except, input validation, stricter regex) over restructuring.

4. Test

Run the patched code against real data when possible:

• Scripts: Execute with --scan (dry-run) first if the script supports it. Compare output against known-good results.
• Skill/command markdown: Verify no broken references to ${CLAUDE_PLUGIN_ROOT}, no missing files.
• Structural: Confirm all files referenced in skill SKILL.md or commands exist in the tree.

Validation checklist:

# All plugin files present
find /path/to/copy -type f
# JSON valid
python -m json.tool /path/to/copy/.claude-plugin/plugin.json
# No broken internal refs
grep -r 'CLAUDE_PLUGIN_ROOT' /path/to/copy --include='*.md'

If the script has a --scan mode, run it against the user's target folder and show the plan without executing.

5. Deploy

1. Build the .plugin zip (always via /tmp to avoid permission issues):

   cd /path/to/copy && zip -r /tmp/plugin-name.plugin . -x "*.DS_Store" && cp /tmp/plugin-name.plugin /path/to/outputs/plugin-name.plugin
   

2. Present the .plugin file link to the user. It renders as a rich preview in Cowork.
3. Copy patched script(s) to the canonical skill location:

   cp patched-script.py mnt/Alex/AGENTS.md/skills/<skill-name>/
   

4. Also save a copy to Downloads as a convenience backup.

6. Update Memory

After successful deployment:

• Update the plugin's auto-memory file with patches applied, date, and any remaining issues.
• Remove resolved pending items.

Constraints

• Never modify files in mnt/.remote-plugins/ directly — always copy first.
• Never reformat code that isn't part of the fix.
• Never add linting or style changes.
• If a test fails, stop and report to the user before deploying.
• The .plugin filename must match the name field in plugin.json.

Additional Resources

• references/plugin-structure.md — quick reference for plugin directory layout and packaging