Self-Improving Agent → ExpertPack
Converts a Self-Improving Agent skill's .learnings/ directory (3.8K ClawHub installs) into a properly structured ExpertPack.
Supported sources:
- LEARNINGS.md — corrections, knowledge gaps, best practices, simplify-and-harden patterns
- ERRORS.md — command failures, exceptions, integration issues
- FEATURE_REQUESTS.md — user-requested capabilities and implementation notes
- Promoted content — entries already promoted to CLAUDE.md, AGENTS.md, SOUL.md, TOOLS.md (detected and cross-referenced)
Usage
cd /root/.openclaw/workspace/ExpertPack/skills/self-improving-to-expertpack
python3 scripts/convert.py \
--workspace /path/to/your/workspace \
--output ~/expertpacks/my-learnings-pack \
[--name "My Agent's Learnings"] \
[--type auto|person|agent|process]
Override .learnings/ location with --learnings /path/to/.learnings.
What It Produces
A complete ExpertPack conforming to schema 2.3:
manifest.yaml(with context tiers, EK stub)overview.mdsummarizing conversion (entry counts, categories, priority breakdown)- Structured directories mapped from learning types:
mind/— best practices, conventions, behavioral patterns, promoted rulesfacts/— knowledge gaps filled, project-specific factsoperational/— error resolutions, tool gotchas, integration fixessummaries/— pattern analyses, recurring issue summariesrelationships/— cross-references between related entries
_index.mdfiles, lead summaries,glossary.md(if terms/tags found)relations.yaml(from See Also links and shared tags)- Clean deduplication preferring promoted > resolved > pending entries
Secrets are automatically stripped (sk-, ghp_, tokens, passwords). Warnings emitted for any found.
Post-Conversion Steps
cd ~/expertpacks/my-learnings-pack- Verify content files are 400–800 tokens each (Schema 2.5 — retrieval-ready by design)
- Measure EK ratio:
python3 /path/to/expertpack/tools/eval-ek.py . - Review
overview.mdandmanifest.yaml - Commit to git and publish to ClawHub
Learn more: https://expertpack.ai • ClawHub expertpack skill
See also: Self-Improving Agent skill on ClawHub.