Step 1 — Identify targets

This is an execution-aware skill — check skills/compact-learnings.learnings.md in the vault before proceeding. If it exists, apply accumulated learnings.

The user has invoked /workbench:compact-learnings, or the skills protocol flagged a learnings file above the 30-entry threshold.

Step 1 — Identify targets

If the user provided a skill name as an argument, target only skills/{name}.learnings.md.

Otherwise, scan all learnings files (default path shown — resolve via config):

find ~/Documents/Claude/Memory/skills -name "*.learnings.md" 2>/dev/null

For threshold-triggered runs, process only the file that triggered it.

Step 2 — For each learnings file

Read the file. Count entries (## headings = entries).

If under 30 entries and this is an unprompted manual run (no specific skill), ask: "Only {N} entries — compact anyway?"

Classify the skill

Determine if this is a workbench plugin skill by searching for a matching SKILL.md in installed plugins whose directory name includes claude-workbench:

find ~/.claude/plugins/installed/*claude-workbench*/skills -name "SKILL.md" -path "*/{skill-name}/*" 2>/dev/null

• Workbench plugin skill → hybrid mode (compact + offer integration into installed SKILL.md, then sync to git repo)
• Any other skill → compact only (rewrite learnings, don't touch SKILL.md)

Walk through each entry

Present each learning to the user with three options and your recommendation.

For workbench plugin skills:

Option	Meaning
Integrate	Bake into SKILL.md — improves the skill definition permanently
Keep	Retain in compacted learnings — relevant but too environment-specific for the definition
Drop	Stale, contradicted, or no longer relevant — remove

For all other skills:

Option	Meaning
Keep	Retain in compacted learnings
Rewrite	Valid learning but poorly worded — rewrite concisely
Drop	Remove

For each entry, state your recommendation and a one-line rationale. Wait for the user's choice before moving on.

Step 3 — Apply changes

Compact the learnings file

Rewrite with only kept/rewritten entries. Maintain chronological order. Write via mcp__plugin_workbench-core_memory__write (full overwrite).

If all entries were dropped or integrated, write a minimal file with just frontmatter and no entries.

Integrate into SKILL.md (workbench plugin skills only)

For entries marked "Integrate":

1. Read the current SKILL.md from the installed plugin directory found in Step 2.
2. Determine where each learning fits — it may modify an existing step, add a caveat, or add a note.
3. Present the proposed changes to the user for approval before writing.
4. Write the updated SKILL.md to the installed copy.
5. Sync to the git source repo: read the repository field from the plugin's plugin.json to identify the GitHub repo. Derive the local clone path by finding a directory whose git remote -v origin matches the repository URL. Copy the updated SKILL.md to the corresponding path in the source repo.

Weave integrated learnings into the existing structure. Do NOT append a "learnings" section — the guidance should read as if it was always part of the skill.

If multiple learnings point to the same issue, consolidate into a single change.

Step 4 — Report

compact-learnings: {skill-name}
  entries: {total} → integrated: {n}, kept: {n}, dropped: {n}
  SKILL.md: {updated|unchanged}

Notes

• Skip files with zero entries silently.
• The 30-entry threshold is a guideline — the user can force a run at any count.
• Multiple files due: If scanning all learnings and more than one file is above threshold, present a summary table first (skill name, entry count, workbench or not) and let the user pick which to process this run. Don't force a 150-decision marathon.
• Preserve the SKILL.md's voice and structure when integrating.