Day 3: OpenClaw SOUL.md, USER.md & AGENTS.md — Give Your AI Assistant a Soul (2026)
2026/05/23

Day 3: OpenClaw SOUL.md, USER.md & AGENTS.md — Give Your AI Assistant a Soul (2026)

OpenClaw Day 3 tutorial: define AI assistant personality with SOUL.md, user context in USER.md, and behavior rules in AGENTS.md — the soul trio for private agents on One Claw or self-host.

Part of the 7-day OpenClaw learning path on One Claw — SEO-friendly tutorials for hosted OpenClaw, private AI assistant, and OpenClaw skills.

Day 3 of the OpenClaw 7-day course is about identity: the soul trio — SOUL.md, USER.md, and AGENTS.md — so your private AI assistant sounds like yours, not a default ChatGPT voice.

One Claw

Why the soul trio matters

Without these files, every session drifts back to generic helpfulness. With them, OpenClaw can:

  • Match your tone (direct, warm, sarcastic, formal)
  • Respect boundaries (what it must never do)
  • Remember who you are and how agents should behave

This is core OpenClaw tutorial material for both self-hosted gateways and One Claw workspaces (edit via files on the instance or your deployment workflow).

SOUL.md — who the assistant is

SOUL.md defines the agent's character:

  • Name, metaphor, or persona ("cyber cat roommate", "ops copilot")
  • Voice: sentence length, emoji policy, languages
  • Values: proactive vs reactive, risk tolerance

Starter template:

# Soul
You are [Name], a private AI assistant on OpenClaw.
Tone: concise, friendly, no fluff.
You proactively suggest next steps but never send external messages without confirmation.

USER.md — who you are

USER.md is the user profile the agent should internalize:

  • Role, timezone, working hours
  • Projects, domains, naming conventions
  • Pet peeves ("never schedule meetings before 9am")

Update when your life changes — the assistant should not guess wrong about your context.

AGENTS.md — how agents behave

AGENTS.md sets operating rules for tools and sub-agents:

  • Which tools are allowed by default
  • Escalation: when to ask vs act
  • File paths, git conventions, brand terms
  • Safety: no destructive shell without explicit approval

Think of it as runbook + guardrails, not poetry.

How the three files work together

FileQuestion it answers
SOUL.mdHow does it speak and feel?
USER.mdWhom is it helping?
AGENTS.mdWhat is it allowed to do?

On heartbeat and cron runs (Day 6), the agent reloads this context so proactive messages stay on-brand.

Practical exercise (30 minutes)

  1. Copy templates into your OpenClaw workspace root (or sync from your repo).
  2. Fill USER.md with 10 bullets of real context.
  3. Add 5 never-do rules to AGENTS.md.
  4. Send the same prompt before/after editing SOUL.md — compare tone.

On One Claw, treat these files like product config: version them in git if you self-sync, or document changes when multiple teammates share one instance.

Common mistakes

  • Novel-length SOUL.md — models skim; keep scannable headings
  • Stale USER.md — wrong timezone ruins scheduling skills (Day 4)
  • Vague AGENTS.md — "be careful" is useless; specify commands and paths

Day 3 checklist

  • SOUL.md drafted and tested in chat
  • USER.md includes timezone + current projects
  • AGENTS.md lists at least 5 explicit rules
  • One message thread shows noticeably better fit

Key takeaways

  • SOUL + USER + AGENTS = personality, context, policy
  • Invest here before piling on skills (Day 4–5)
  • Tomorrow: connect Gmail, calendar, search, browser so the assistant can do work

Previous & next

  • ← Day 1: Meet OpenClaw
  • ← Day 2: 10-minute setup
  • Day 4: Gmail, calendar & search skills →

One Claw resources

  • OpenClaw setup tutorial · How to start with One Claw · Hosted vs self-hosted

Live demo · Pricing

All Posts

Author

avatar for One Claw Team
One Claw Team

Categories

  • Learn
Why the soul trio mattersSOUL.md — who the assistant isUSER.md — who you areAGENTS.md — how agents behaveHow the three files work togetherPractical exercise (30 minutes)Common mistakesDay 3 checklistKey takeawaysPrevious & nextOne Claw resources

More Posts

Hermes Agent vs Claude Code vs OpenClaw (2026): Why Buyers Are Choosing a Hosted Workspace
NewsProduct

Hermes Agent vs Claude Code vs OpenClaw (2026): Why Buyers Are Choosing a Hosted Workspace

A practical 2026 comparison of Hermes Agent, Claude Code, and OpenClaw — three AI agent philosophies, real pricing, and why a managed hosted workspace usually wins on time-to-value and subscription conversion.

avatar for One Claw Team
One Claw Team
2026/05/19
Customer Support with One Claw: Faster, Steadier Replies—Not Replacing Humans
NewsProduct

Customer Support with One Claw: Faster, Steadier Replies—Not Replacing Humans

From pre-sales Q&A and FAQ drafts to renewal outreach—help your team respond faster and more consistently.

avatar for One Claw Team
One Claw Team
2026/05/08
Day 6: OpenClaw Heartbeat, Cron Jobs & Memory — Proactive AI Assistant (2026)
Learn

Day 6: OpenClaw Heartbeat, Cron Jobs & Memory — Proactive AI Assistant (2026)

OpenClaw Day 6: configure heartbeat checks, cron scheduled tasks, and three-layer memory (daily logs, MEMORY.md, SOUL.md) so your private AI assistant works while you sleep.

avatar for One Claw Team
One Claw Team
2026/05/26

Newsletter

Waitlist

Subscribe to our newsletter for the latest news and updates

One Claw

Hosted OpenClaw in the cloud—subscribe or buy credits, chat, tasks, remote channels, and skills in one workspace.

Support: [email protected]

Product
  • Features
  • Pricing
  • FAQ
Resources
  • Blog
Company
  • About
  • Contact
  • Waitlist
Legal
  • Cookie Policy
  • Privacy Policy
  • Terms of Service
  • Acceptable Use Policy
© 2026 One Claw. All Rights Reserved.
One Claw
  • Features
  • Pricing
  • Blog
  • Waitlist
  • Contact