OpenClaw Proper Usage
Core Rule
- Clarify objective + output format
- Pick one best tool/skill
- Execute in scoped steps
- Verify with concrete evidence
- Report result + next action
Routing
- Planning/strategy/analysis: gemini
- Web search: gemini
- Code review: gemini
- Code implementation/tests: Claude CLI (daedalus-code)
Subagent Pattern
- Split into 2–5 independent streams
- Strict scope + expected output per stream
- Launch, monitor, merge evidence-backed outputs only
- Resolve conflicts explicitly
Commands
sessions_spawn task:"<task+deliverable+constraints>" label:"" model:"azure-openai/gpt-4.1" thinking:"low" agents_list
Completion Contract
Every task response must include: what changed, what ran, pass/fail, risks, next action.
Triage
- Tiny: do it directly
- Ambiguous: clarify first
- Web-heavy: Gemini first
- Planning-heavy: Gemini first, Claude for implementation
- Broad: split → parallelize → merge
- Risky: plan + verify before edits