ExpertPack

Structured knowledge packs for AI agents. Maximize the knowledge your AI is missing.

Learn more: expertpack.ai · GitHub · Schema docs

Full schemas: /path/to/ExpertPack/schemas/ in the repo (core.md, person.md, product.md, process.md, composite.md, eval.md)

Pack Location

Default directory: ~/expertpacks/. Check there first, fall back to current workspace. Users can override by specifying a path.

Actions

1. Load / Consume a Pack

1. Read manifest.yaml — identify type, version, context tiers
2. Read overview.md — understand what the pack covers
3. Load all Tier 1 (always) files into session context
4. For queries: search Tier 2 (searchable) files via RAG or _index.md navigation
5. Load Tier 3 (on-demand) only on explicit request (verbatim transcripts, training data)

OpenClaw RAG config — add to openclaw.json:

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "extraPaths": ["path/to/pack/.chunks"],
        "chunking": { "tokens": 500, "overlap": 0 },
        "query": {
          "hybrid": {
            "enabled": true,
            "mmr": { "enabled": true, "lambda": 0.7 },
            "temporalDecay": { "enabled": false }
          }
        }
      }
    }
  }
}

For detailed platform integration (Cursor, Claude Code, custom APIs, direct context window): read {skill_dir}/references/consumption.md.

2. Create / Hydrate a Pack

1. Determine pack type: person, product, process, or composite
2. Read {skill_dir}/references/schemas.md for structural requirements
3. Scaffold the directory structure per the type schema
4. Create manifest.yaml and overview.md (both required)
5. Populate content using EK-aware hydration:
   • Blind-probe each extracted fact before filing
   • Full treatment for EK content (the model can't produce it)
   • Compressed scaffolding for GK content (the model already knows it)
   • Skip content with zero EK value
6. Add retrieval layers: _index.md per directory, summaries/, propositions/, glossary.md
7. Add sources/_coverage.md documenting what was researched

For full hydration methodology, EK triage process, and source prioritization: read {skill_dir}/references/hydration.md.

3. Chunk for RAG

Run the schema-aware chunker:

python3 {skill_dir}/scripts/chunk.py --pack <pack-path> --output <pack-path>/.chunks

• Respects ## headers, lead summaries, proposition groups, <!-- refresh --> metadata
• Each output file = one semantically coherent chunk
• Point OpenClaw RAG at .chunks/ with overlap=0

Why this matters: Schema-aware chunking produced +9.4% correctness and -52% tokens vs. generic chunking in controlled experiments. It's the single highest-impact consumption optimization.

4. Measure EK Ratio & Run Quality Evals

For EK ratio measurement (blind probing) and automated quality evals, install the companion skill:

clawhub install expertpack-eval

See expertpack-eval for full details on EK measurement, eval runner, and the improvement loop.

5. Backup / Export OpenClaw → ExpertPack

Export an OpenClaw agent's accumulated knowledge into a structured ExpertPack composite.

Step 1 — Scan:

python3 {skill_dir}/scripts/scan.py --workspace <workspace-path> --output /tmp/ep-scan.json

Review the scan output with the user. It proposes pack assignments (agent, person, product, process) with confidence scores. Flag ambiguous classifications for user decision.

Step 2 — Distill (repeat per pack):

python3 {skill_dir}/scripts/distill.py \
  --scan /tmp/ep-scan.json \
  --pack <type:slug> \
  --output <export-dir>/packs/<slug>/

• Distill, don't copy — target 10-20% volume of raw state
• Strips secrets automatically (API keys, tokens, passwords)
• Deduplicates, prefers newest for conflicts

Step 3 — Compose:

python3 {skill_dir}/scripts/compose.py \
  --scan /tmp/ep-scan.json \
  --export-dir <export-dir>/

Generates composite manifest and overview.

Step 4 — Validate:

python3 {skill_dir}/scripts/validate.py --export-dir <export-dir>/

Checks: required files exist, manifest fields valid, no secrets leaked, file sizes within guidelines, cross-references resolve.

Step 5 — Review & ship. Present validation report to user. They decide whether to commit/push.

Critical rules:

• Never include secrets in the export
• Never modify the live workspace — all output goes to the export directory
• Flag personal information for access tier review
• Default user-specific content to private access