ExpertPack

Work with ExpertPacks — structured knowledge packs for AI agents. Use when: (1) Loading/consuming an ExpertPack as agent context, (2) Creating or hydrating a...

Security Scan

The skill's requirements, instructions, and bundled scripts are coherent with its stated purpose (creating, chunking, scanning, and exporting ExpertPacks); it scans and manipulates local workspace files but does not request unrelated credentials or perform obvious network exfiltration.

Purpose & Capability

Name/description (ExpertPack: create, hydrate, chunk, export packs) aligns with the provided scripts and SKILL.md. Required binary is only python3, which is appropriate for the included Python tools. No unrelated environment variables, binaries, or external services are requested.

Instruction Scope

SKILL.md explicitly instructs the agent to scan the user's workspace, run chunk/distill/compose/validate/scan scripts, and export to a separate directory. This is within the declared purpose, but the scan reads many local files (memory/, logs, scripts, etc.), so it legitimately accesses potentially sensitive data. The README and scripts also state 'Never include secrets in the export' and implement redaction heuristics, but redaction is heuristic and not a foolproof substitute for user review.

Install Mechanism

No install spec or external downloads — instruction-only with bundled Python scripts. The scripts avoid external package installs (they fall back to builtin parsing when PyYAML is absent). There are no downloads from arbitrary URLs or extracted archives in the manifest.

Credentials

The skill declares no required environment variables or primary credentials, which matches its local-file-focused behavior. However, the scripts read configuration files (openclaw.json) and arbitrary workspace files; the distill/scan scripts include secret-detection and redaction regexes. This is appropriate but imperfect — redaction may miss secrets or sensitive PII, so review is required before exporting or sharing outputs.

Persistence & Privilege

always:false and default autonomous invocation are present (the platform default). The skill writes only to the specified export directory and scaffolds pack files there; it does not declare system-wide persistence or modify other skills' configs. SKILL.md explicitly instructs not to modify the live workspace.

Assessment

This skill is coherent for exporting and preparing ExpertPacks, but it scans your workspace and reads many local files (memory journals, logs, scripts, config). Before running: (1) run the scan step and carefully review the generated /tmp/ep-scan.json to confirm what will be exported, (2) choose an explicit export directory you control (do not use a system or shared directory), (3) review the distilled/scaffolded output before sharing it externally — the redaction is heuristic and can miss secrets or PII, (4) if you have highly sensitive data, run the tools in an isolated environment (local VM/container) or remove sensitive files from the workspace first, and (5) optionally inspect the included scripts yourself (they are plain Python) to validate there are no unexpected network calls before use.

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