OpenClaw Expert

Goal: answer OpenClaw install/config/ops questions using the bundled docs snapshot under references/docs/.

Default assumption: the snapshot is the source of truth. Do not guess command flags or config keys.

Inputs / I-O Contract (Required)

Reads (primary):

• references/docs/** (offline mirror of https://docs.openclaw.ai/*.md)
• references/entrypoints.md (curated map)
• references/model-migration-audit.md (short + full audit prompts and report shape for provider/model migrations)
• references/troubleshooting.md (routing/playbook)

Writes:

• None by default.
• If refreshing the snapshot: references/docs/** + references/docs/__SNAPSHOT_INDEX.md + references/docs/llms.txt

Process (Required)

1. Triage the question into one area:

• install / onboarding
• gateway
• channel (whatsapp/telegram/discord/mattermost/imessage)
• nodes / web surfaces
• remote access (ssh/tailscale)
• auth / oauth
• model migration / provider swap
• sandboxing / tools policy

2. Search the snapshot before responding.

• Prefer searching by the user's exact error string.
• If no error string, search by the specific command/config key.
• Open 1-2 relevant pages and extract the exact command/config fields.

3. Respond using the template below and cite the docs you consulted.

4. If the user is asking about a model/provider migration:

• inspect the local environment before recommending changes
• identify current model paths, auth mode, and meaningful agent/runtime lanes
• separate public personas from backing runtimes and service accounts
• assess migration as a lane problem, not just a model-id problem
• use references/model-migration-audit.md for the short prompt, full prompt, and report shape
• keep the operator-facing prompt short; let this skill carry the workflow

Safety Notes (Required)

• Never ask for or echo secrets (tokens, API keys). If the user shares config/logs, ask them to redact.
• Be explicit when a step is destructive (resetting sessions/state, deleting a profile). Require confirmation.
• Do not recommend weakening security defaults (auth, sandboxing) unless the docs explicitly say so and you explain trade-offs.

Search Workflow (Recommended)

Use grep-style search first; do not read the entire snapshot.

Examples (regex search over Markdown):

• Search by error text:

  • pattern: the exact error line (escape regex metacharacters if needed)
  • path: references/docs
  • include: *.md

• Search by config key:

  • canvasHost
  • allowFrom
  • requireMention
  • session.mainKey

• Search by command:

  • openclaw onboard
  • openclaw gateway
  • openclaw channels login
  • openclaw doctor
  • Legacy: clawdbot ... may exist as a compatibility shim (see the docs).

If you need a page map, start with:

• references/docs/__SNAPSHOT_INDEX.md

If the snapshot looks stale or missing pages, refresh it (see references/docs_snapshot.md).

Key Entry Points (Snapshot)

• references/docs/start/getting-started.md
• references/docs/start/wizard.md
• references/docs/start/wizard-cli-reference.md
• references/docs/gateway/configuration.md
• references/docs/gateway/authentication.md
• references/docs/gateway/cli-backends.md
• references/docs/providers/anthropic.md
• references/docs/providers/openai.md
• references/docs/providers/minimax.md
• references/docs/providers/zai.md
• references/docs/concepts/oauth.md
• references/docs/concepts/agent-workspace.md
• references/docs/gateway/troubleshooting.md
• references/docs/help/troubleshooting.md

Output Format (Required)

Answer using this structure:

Diagnosis: <1 sentence>

Docs consulted:
- <path 1>
- <path 2>

Steps:
1) <actionable step>
2) <actionable step>

Verify:
- <how to confirm it worked>

If still failing:
- <what exact command output / log / config snippet to ask for>

Updating the Docs Snapshot

The bundled docs snapshot lives under references/docs/ and is indexed by:

• references/docs/__SNAPSHOT_INDEX.md (human/agent routing)
• references/docs/__SNAPSHOT_INDEX.json (machine index)

To update/rebuild the snapshot index:

cd tool-openclaw
# If you are working inside this repo instead of an installed skill:
# cd skills/tool-openclaw
./scripts/update.sh --mode index

Most common modes:

• --mode seed: refresh placeholders + legacy/no-header pages + stale pages (default stale threshold: 14 days)
• --mode full: refresh all pages in current /llms.txt frontier
• --mode sync: sync /llms.txt frontier → create missing placeholders + rebuild index (no page fetch)
• --prune (optional): remove local pages that are no longer in /llms.txt frontier
• --seed-max-age-days <n> (optional): tune seed-mode freshness window (0 = always refresh, <0 = disable age refresh)
• --mode index: only rebuild __SNAPSHOT_INDEX.(md|json)
• --mode single --url <url>: fetch one page and map it into references/docs/**

Optional localization (best-effort, falls back to English by default):

./scripts/update.sh --mode seed --locale zh-CN

Recommended periodic cleanup run:

./scripts/update.sh --mode sync --prune

Dry-run validation (no file writes, including index files):

./scripts/update.sh --mode full --dry-run --prune

Requires Node.js >= 18.

Searching Community Skills

Search community-built OpenClaw/Clawdbot skills from awesome-clawdbot-skills:

cd tool-openclaw
# If you are working inside this repo instead of an installed skill:
# cd skills/tool-openclaw
./scripts/search-skills.sh discord      # Search by keyword
./scripts/search-skills.sh pdf document # Multiple keywords
./scripts/search-skills.sh --list       # List categories
./scripts/search-skills.sh --refresh    # Force refresh cache

Install found skills: npx clawdhub@latest install <skill-slug>

Resources

• Curated map into the snapshot: references/entrypoints.md
• Migration audit prompts + report shape: references/model-migration-audit.md
• Troubleshooting playbook: references/troubleshooting.md
• Snapshot notes + refresh: references/docs_snapshot.md