Agent Init Skill for Claude

Agent Init

Initialize OpenClaw agent workspace with tailored MD files through an interactive interview.

Workflow

Phase 1: Interview (MANDATORY — do not skip)

Before generating ANY files, gather context through conversation. Ask in batches of 2-3 questions:

Batch 1 — Identity & Purpose:

• What is this agent's primary purpose? (e.g., coding assistant, research, DevOps, personal assistant)
• What name and emoji? Any personality traits?

Batch 2 — User Profile:

• Who will use this agent? (name, timezone, preferences)
• Communication style preference? (formal/casual, verbose/terse, language)

Batch 3 — Environment:

• Is this a container instance or external (host) instance?
• What tools/languages does the agent need? (Python, Node, etc.)
• Any specific workflows or periodic tasks?

Batch 4 — Boundaries:

• Any topics or actions the agent should avoid?
• Privacy requirements beyond defaults?

Only proceed to Phase 2 after user confirms the interview is complete.

Phase 2: Environment Check

Run scripts/check-env.sh to detect Python/uv status:

bash <skill-path>/scripts/check-env.sh

If uv is missing and user wants Python support:

bash <skill-path>/scripts/check-env.sh --install

For container instances, run inside the container:

docker exec <containerId> bash -c "which uv && uv --version || echo 'uv: NOT FOUND'"

Phase 3: Generate Files

Generate files in this order, showing each to user for confirmation before writing:

• IDENTITY.md — Fill in fields from interview
• USER.md — Fill in user profile
• SOUL.md — Rewrite content, keep 4-section structure (Core Truths / Boundaries / Vibe / Continuity)
• TOOLS.md — Add environment info, Python/uv config
• AGENTS.md — Extend default template with domain-specific sections (see strategy below)
• HEARTBEAT.md — Add periodic tasks if any
• BOOTSTRAP.md — Skip unless user wants first-run ritual

AGENTS.md Strategy: Extend, Don't Replace

Read the current AGENTS.md first. The default template contains critical infrastructure:

• Session startup sequence (file loading order)
• Memory system (daily + MEMORY.md)
• Safety rules
• Heartbeat logic

Add new sections; never remove existing ones. Safe insertion points:

• After "Every Session" → domain-specific startup tasks
• After "Safety" → additional security rules
• After "Tools" → Python/uv preferences
• Before "Make It Yours" → project-specific workflows

Mandatory TOOLS.md Additions

## Python
- Package manager: `uv` (NEVER use pip directly)
- Create venv: `uv venv .venv`
- Install: `uv pip install <package>`
- Run: `uv run python script.py`
- If uv missing: `curl -LsSf https://astral.sh/uv/install.sh | sh`

Phase 4: Write Files

Determine write method based on instance type:

External instance (workspace on host filesystem):

# Workspace files live at {workspacePath}/workspace/ (or workspace-{profile}/)
cat > {workspacePath}/workspace/SOUL.md << 'ENDOFFILE'
[content]
ENDOFFILE

Container instance (OpenClaw running in Docker):

docker exec -i <containerId> sh -c 'cat > /home/node/.openclaw/workspace/SOUL.md' << 'ENDOFFILE'
[content]
ENDOFFILE

For non-main agents (agentId ≠ "main"):

• Check if agent has a dedicated workspace via config.get
• Agent workspace might be at workspace-{agentId}/ or configured separately

Phase 5: Verify

After writing, confirm files are in place:

# External
ls -la {workspacePath}/workspace/*.md

# Container
docker exec <containerId> ls -la /home/node/.openclaw/workspace/*.md

Reference Files

• references/templates.md — Official templates, loading order, per-file strategy, section structure
• references/openclaw-workspace.md — Instance types, workspace paths, Python env, agent creation rules

Read these when you need detailed guidance on template structure or workspace configuration.

Rules

• Interview first — never generate files without understanding the user's intent
• Extend, don't replace — AGENTS.md default template is infrastructure, not boilerplate
• Show before write — display each generated file for user confirmation
• uv over pip — always configure uv as the Python package manager
• No secrets in files — workspace files are injected into every prompt turn
• Keep files concise — all workspace files consume tokens every turn (20KB/file limit, 150KB total)
• Respect existing content — read before write, merge non-destructively