Prompt Guide

Everything in JARVIS is a prompt. The system prompt, the agent personalities, the skill instructions, the tone registers. They're all just files you edit.

The 4-layer prompt architecture

LAYER 1   System (SOUL.md)         →  Permanent identity, voice, rules
LAYER 2   Agent SOUL                →  Per-agent specialization
LAYER 3   Skill SKILL.md            →  Per-task playbook
LAYER 4   Memory + session context  →  Recent conversations, preferences

Editing rule of thumb

Want JARVIS to behave differently every time → edit SOUL.md
Want one agent to behave differently → edit agents/<id>/SOUL.md
Want a specific TASK to follow a precise procedure → write a skill
Want JARVIS to remember a fact → put it in MEMORY.md

Editing SOUL.md

Your main system prompt lives at ~/.openclaw/workspace/SOUL.md.

Best practices

Do	Don't
Use specific examples	Vague abstractions like "be witty"
Include negative examples ("Never say X")	Only positive examples
State preferences as identity	Indirect pleas ("please be direct")
Critical rules at TOP and BOTTOM	Rules buried in the middle
Test edge cases by Telegram-messaging	Assume the prompt works without testing

Tone Registers

The kit ships with 5 tone registers. The model picks one per turn based on context.

BUTLER, default, low-stakes, briefings
TACTICAL, trading alerts, position events, capital moves
DIAGNOSTIC, system incidents, failures
OPERATOR, executing direct requests
CONFIDANT, venting, frustrated, late-night messages

Anti-patterns

Conflicting instructions. Pick one stance and commit.
Negative-only instructions. Add what TO do alongside what NOT to do.
Telling the model what it is. Strips personality. Replace with role-as-action.
Apology lists. Trains the model to hedge constantly.
10-page SOUL.md. Cap at 3-5 pages. Push specifics to skills or memory.

Customizing the 14 agents

Each agent has its own SOUL at ~/.openclaw/agents/<id>/agent/SOUL.md.

Adding a new agent

mkdir -p ~/.openclaw/agents/medic/agent
cat > ~/.openclaw/agents/medic/agent/SOUL.md <<'EOF'
# SOUL.md, Medic
You are Medic. The on-call doctor for the system.
Triage, stabilize, escalate only if needed.
...
EOF
# Then register in openclaw.json → agents.list[]

Hot-reloading

Change	How JARVIS picks it up
Edit SOUL.md / MEMORY.md / agent SOULs	Next turn
Edit a skill	Next time the skill is invoked
Edit `openclaw.json`	`openclaw doctor --fix`

Iteration loop

Use it for a week. Note moments where JARVIS felt off.
Categorize. Was it voice? Decision? Missing context? Wrong priority?
Edit the right layer. Voice issue → SOUL. Skill broke → that skill. Forgot a fact → MEMORY.
Test the edit. Send 5 messages that exercise the change. Iterate.