OpenClaw Advanced Guide: 5 Modules to Turn Your AI Assistant from Basic to Expert

Prerequisites

OpenClaw must be installed, at least one channel connected, the configuration file ~/.openclaw/openclaw.json must exist, and an AI model (Claude, GPT, etc.) must be configured. Verify the basic setup with openclaw doctor --fix.

1. Identity System

Three markdown files in ~/.openclaw/workspace define the agent’s persona.

SOUL.md – core values, behavior rules, and boundaries. Example content includes "Core Truths" such as "Be genuinely helpful" and "Have opinions", and "Boundaries" such as "Private things stay private".

IDENTITY.md – name, avatar, emoji, and vibe. Example content shows Name: Claw, Creature: AI familiar, Vibe: Sharp but warm, Emoji: 🐾, Avatar: avatars/openclaw.png.

USER.md – user preferences, project context, and notes. Example content includes name, pronouns, timezone, notes, and a context block with current project, tech stack, and experience level.

At minimum SOUL.md and IDENTITY.md give the agent a usable persona; USER.md can be added incrementally.

2. Persistent Memory

Workspace layout:

~/.openclaw/workspace/

├── MEMORY.md          # long‑term memory (main session only)

├── memory/

│   ├── 2026-03-01.md  # daily log

│   ├── 2026-02-28.md

│   └── 2026-02-27.md

├── SOUL.md

├── IDENTITY.md

└── USER.md

Two memory tools are provided: memory_search – semantic vector search (default provider OpenAI text-embedding-3-small). memory_get – read a specific file or line range.

Hybrid search configuration (vector weight 0.7, BM25 weight 0.3, MMR re‑ranking enabled, temporal decay with 30‑day half‑life) can be set in openclaw.json as shown.

When the session nears the token limit, the MemoryFlush mechanism automatically writes important notes to MEMORY.md using a silent NO_REPLY prompt. Example configuration includes reserveTokensFloor, softThresholdTokens, and custom system prompts.

3. Skills – Extending Agent Capabilities

Skills are plug‑ins that let the agent call APIs, manipulate files, or integrate third‑party services. Load order (high to low priority): <workspace>/skills (personal), ~/.openclaw/skills (shared), bundled skills (built‑in), and directories listed in skills.load.extraDirs. When two skills share the same name, the higher‑priority version overrides the lower one.

Skill definition uses a YAML front‑matter block. Example SKILL.md for a Gemini image generator includes fields name, description, and metadata.openclaw (requires binaries, environment variables, config, primaryEnv, emoji, homepage).

Enable a skill in openclaw.json under skills.entries with enabled: true and optional apiKey, env, and config sections, as illustrated.

The official skill marketplace ClawHub works like npm or PyPI. Common commands:

# Install a skill
clawhub install github-integration
# Update all skills
clawhub update --all
# Search for a skill
clawhub search "notion"

Custom skills can be placed in ~/.openclaw/skills/your-skill/SKILL.md and enabled via the JSON config. A minimal todo‑manager skill example demonstrates required fields and usage instructions.

4. Sub‑Agents – Parallel Processing

Sub‑agents run in the background, keeping the main session responsive. Each sub‑agent has an identifier like agent:<id>:subagent:<uuid>, guaranteeing isolation.

Asynchronous execution prevents long‑running tasks from blocking the main conversation.

Supports nested orchestration (depth‑2 sub‑agents) for complex workflows.

Concurrency limits ( maxConcurrent, maxChildrenPerAgent) prevent resource exhaustion.

Invocation uses a JSON payload. Example:

{
  "task": "Analyze the latest logs for errors",
  "label": "log-analysis",
  "model": "anthropic/claude-sonnet-4-5",
  "thinking": "medium",
  "thread": false,
  "mode": "run"
}

An orchestrated multi‑agent code‑review example shows three sub‑agents, each reviewing a PR with a different model and depth.

Slash commands manage sub‑agents, e.g., /subagents list, /subagents kill <id>, /subagents spawn reviewer-1 "Analyze error logs".

5. Cron Jobs – Automated Tasks

Cron jobs let the agent perform scheduled actions. Three schedule kinds are supported: at – a single ISO‑8601 timestamp. every – fixed millisecond interval. cron – standard 5‑field cron expression.

Example one‑off reminder (system event):

{
  "name": "Reminder",
  "schedule": {"kind": "at", "at": "2026-03-01T16:00:00Z"},
  "sessionTarget": "main",
  "wakeMode": "now",
  "payload": {"kind": "systemEvent", "text": "Reminder text"}
}

Example recurring daily brief (announce to WhatsApp):

{
  "name": "Morning brief",
  "schedule": {"kind": "cron", "expr": "0 7 * * *", "tz": "America/Los_Angeles"},
  "sessionTarget": "isolated",
  "payload": {"kind": "agentTurn", "message": "Summarize overnight updates."},
  "delivery": {"mode": "announce", "channel": "whatsapp", "to": "channel:C1234567890"}
}

CLI helpers:

# Add a one‑off reminder
openclaw cron add \
  --name "Reminder" \
  --at "2026-03-15T18:00:00Z" \
  --session main \
  --system-event "Submit expense report." \
  --wake now

# Add a recurring daily summary
openclaw cron add \
  --name "Daily summary" \
  --cron "0 17 * * 1-5" \
  --tz "America/New_York" \
  --session isolated \
  --message "Review today’s memory/logs and summarize key accomplishments." \
  --announce \
  --channel telegram

# List, run, or delete jobs
openclaw cron list
openclaw cron run job-abc123
openclaw cron delete job-abc123

If a job fails, OpenClaw retries up to three times and logs details under ~/.openclaw/cron/runs/. Timeout defaults to 15 minutes but can be changed in the JSON config.

6. Next‑Step Recommendations

Suggested learning path:

Master the identity system (create SOUL.md, IDENTITY.md, USER.md).

Configure the memory layer (vector search, experiment with hybrid settings).

Install 2‑3 skills from ClawHub and develop a simple custom skill.

Enable sub‑agents for parallel workloads and set up cron jobs for routine automation.

Explore orchestrator mode for multi‑depth task trees.