SOUL.md Explained: Set Your OpenClaw Agent’s Personality, Tone, and Boundaries

Most people don’t need a “smarter” AI.

They need an AI that stops being weird.

That is what SOUL.md is for.

If OpenClaw is your digital employee:
- AGENTS.md is the employee handbook.
- SOUL.md is the culture and brand voice guide.

SOUL.md tells your agent:
- how to sound
- what it must never do
- how to format outputs
- when to ask you for approval

If you are using OpenClaw for customer communication, SOUL.md is one of the most important risk controls you have.

Because it is how you prevent:
- overpromising
- snarky tone
- long rambly replies
- accidental “sure, I sent it” behavior

---

0) What this guide will help you do

By the end of this guide, you’ll be able to:
- write a SOUL.md that makes your agent feel like a real assistant (not a random chatbot)
- set hard boundaries that reduce business risk
- standardize formatting so outputs are easy to approve
- test SOUL.md quickly and tighten it without overthinking

---

1) What SOUL.md does (and why owners should care)

Officially, SOUL.md defines the agent’s persona, tone, and boundaries, and it is loaded every session.

Plain English:
- Every time you talk to the agent, SOUL.md is part of what it “sees.”
- So the agent does not drift into random styles.

SOUL.md is not a “cute personality prompt.”

It’s an operational document.

It’s what makes the difference between:
- “This sounds like us.”

and

• “Why is it talking like a 22-year-old influencer?”

The SMB reason this matters

In a small business, you do not have layers of reviewers.

Your agent’s first draft often becomes:
- the email you send
- the DM you paste
- the proposal you forward

So the default voice matters.

---

2) What to include (the 5-part framework)

If you are staring at a blank SOUL.md, use this structure.

Part 1: Voice (tone)

How should it sound? - warm and professional - concise - no slang

Part 2: Audience

Who is it talking to? - customers - vendors - internal staff

You can include different rules by audience, but keep it readable.

Part 3: Boundaries (non-negotiables)

This is where you prevent headaches.

Common SMB boundaries:
- draft-first
- approval before sending
- no pricing promises
- no legal/medical/financial advice
- no commitments on timelines unless you provided them

Part 4: Formatting rules

This improves usability.

Examples:
- short paragraphs
- bullet points
- include “Next step” at the end

Part 5: Approval rules

Be explicit.

If the agent might message a customer, you want a clear rule like:
- “Never send externally without approval.”

Here is a simple SOUL skeleton you can copy:

# SOUL.md

## Voice
- Warm, professional, plain English
- Short paragraphs
- No slang

## Boundaries
- Draft first. Never send externally without approval.
- No promises on pricing or timelines.
- No legal/medical/financial advice.

## Formatting
- Start with a 1-line summary.
- Then bullets.
- End with a clear question for approval.

---

3) The most important part: boundaries that keep you safe

If you only write one section well, write this one.

Here are the boundaries that prevent 80% of SMB problems.

Boundary A: Draft-first (always)

Make it explicit:
- the agent drafts
- you approve
- then you send

This prevents:
- accidental customer messages
- “we can do it tomorrow” promises
- weird apologies

Boundary B: No pricing promises

If your pricing changes based on scope, do not allow confident numbers.

Use rules like:
- “Offer a range only if provided in notes.”
- “If uncertain, ask for details and say we’ll confirm pricing.”

Boundary C: No timeline promises

Same concept. Replace “We can definitely do Friday” with:
- “We can likely fit you in Friday—confirming schedule now.”

Boundary D: No legal/medical/financial advice

Even if you’re in those industries, the agent should:
- provide general information
- recommend professional review

Boundary E: Don’t invent policies

This sounds obvious, but it’s a real failure mode.

Add:
- “If you don’t know the policy, ask me. Do not guess.”

---

4) Formatting rules that make approvals fast

Owners don’t hate writing.

They hate:
- messy drafts
- rambling paragraphs
- missing next steps

Formatting rules are how you save time.

Here are high-leverage rules:
- keep replies under 120–180 words unless asked
- use short paragraphs (1–3 lines)
- use bullets for steps
- end with one clear question

“Approval-ready” output pattern

A clean pattern that works almost everywhere:

1) One-line summary
2) Draft text (ready to paste)
3) Notes / questions
4) “Approve to send?”

That structure reduces back-and-forth.

---

5) Before vs After (what SOUL.md fixes)

This is what owners usually feel.

Before (no SOUL)

- Replies are inconsistent. - Sometimes it’s too formal, sometimes too casual. - It over-apologizes or over-explains. - You spend more time editing than writing.

After (clear SOUL)

- Drafts sound like the business. - Replies are short and usable. - The agent asks for approval instead of acting. - You only tweak specifics (names, dates, prices).

---

6) SMB examples: SOUL.md setups that work

These are intentionally practical. Copy/paste and adjust.

Example 1: Home services (plumber / HVAC / electrician)

# SOUL.md

You are a friendly, efficient office assistant for a local home services company.

## Voice
- Warm, calm, plain English
- Direct, scheduling-first
- No slang, no emojis

## Boundaries
- Draft only. Never send externally without approval.
- No pricing promises. If pricing is unknown, ask for photos/details and say we’ll confirm.
- No timeline guarantees. Offer windows, not certainties.
- Always confirm: name, address/zip, problem summary, preferred time window.

## Formatting
- Keep replies under 140 words unless asked.
- Ask max 2 clarifying questions.
- End with: “Want me to send this?”

Example 2: Professional services (accounting / consulting)

# SOUL.md

You are a professional assistant for a small firm.

## Voice
- Professional, confident, not salesy
- Short paragraphs

## Boundaries
- Draft only. Never send without approval.
- No tax/legal advice. Provide general info and recommend a professional review.
- No commitments about outcomes or deadlines unless confirmed.

## Formatting
- Start with a 1-line summary.
- Provide 2 options when drafting (short + slightly longer).
- End with a single approval question.

Example 3: Creator / brand voice

# SOUL.md

You are a creative assistant for a content creator.

## Voice
- Upbeat, punchy, brand-safe
- Simple language; avoid jargon

## Boundaries
- No controversial topics unless explicitly approved.
- Don’t claim results we can’t prove.
- Draft only; no posting without approval.

## Formatting
- Provide 3 variations for hooks.
- Provide 3 CTA options.
- Keep captions under 220 words unless asked.

---

7) Step-by-step: Write your SOUL.md in 15 minutes

This is the fastest way to get a usable version.

Step 1: Write a “voice sentence”

Fill in this blank:
- “We sound ___, ___, and ___. We do not sound ___.”

Examples:
- “We sound calm, professional, and friendly. We do not sound sarcastic or hypey.”
- “We sound direct and helpful. We do not sound overly formal.”

Step 2: Add 3–5 boundaries you will not compromise on

Start here if you’re stuck:
- Draft only, approval required.
- No pricing promises.
- No timeline guarantees.
- No invented policies.
- No legal/medical/financial advice.

Step 3: Add formatting rules

Pick 3 rules:
- max word count
- bullets vs paragraphs
- “end with a question”

Step 4: Add one “clarifying questions” rule

This prevents annoying interrogations.

Example:
- “Ask max 2 clarifying questions unless the user requests a full intake.”

Step 5: Add one “confidence rule”

Example:
- “If you are not sure, say you’re not sure and ask me.”

Step 6: Test and tighten

Run 3 tests (below). Edit only what you disliked.

---

8) How to test and tune SOUL.md in 10 minutes

Testing prevents “surprises” later.

Run these three test prompts:

1) Customer inquiry reply

Draft a reply to this lead:
“Hi, how much for a deep clean this week?”
Remember: draft only, no price promises.

2) Bad review response

Draft a calm public response to this review:
“They were late and the job took forever.”
Keep it respectful. No excuses. Offer a next step.

3) Newsletter or update

Draft a short weekly update for customers.
Tone: warm, professional, plain English.

Then ask yourself:
- Is it too long?
- Does it sound like us?
- Did it overpromise?
- Did it ask for approval?

Make one edit at a time.

---

9) Common mistakes (and fixes)

Mistake: SOUL.md is a novel

Symptom: it gets ignored or truncated in practice.

Fix: keep it short and strict. You can add detail elsewhere.

Mistake: no boundaries

Symptom: risky outputs.

Fix: add 3–5 hard rules that never change.

Mistake: no formatting rules

Symptom: messy drafts that waste your time.

Fix: specify the output format.

Mistake: conflicting instructions

Symptom: the agent flips between “short” and “detailed” randomly.

Fix: be explicit about priority.

Example:
- “If conflict: safety > accuracy > brevity.”

Mistake: trying to encode your whole brand book

Symptom: you spend two hours writing SOUL.md.

Fix: write a v1 in 15 minutes. Tighten later.

---

10) Mini-FAQ

Is SOUL.md the same as a system prompt?
It’s effectively your local “persona + rules” file. It’s loaded every session, so it keeps behavior consistent.

Can I have multiple SOUL files (different voices)?
Yes, but keep one default. If you need multiple voices, use separate agents or explicitly switch context.

Should SOUL.md include pricing and policies?
Only if they are stable and you’re confident they’re correct. Otherwise, add a rule: “Ask me—don’t guess.”

What’s the fastest way to get value?
Write boundaries + formatting rules first. Tone can be refined later.

How do I stop the agent from sounding overly apologetic?
Add a rule:
- “No excessive apologies. One sentence max.”

---

11) Before vs After: messy draft vs approval-ready draft

Here’s a quick illustration of why SOUL.md is worth writing.

Before (no SOUL.md)

> Heyyy! Totally! We can definitely do that this week. Prices vary but it’s usually not too bad. Just send your address and we’ll get you booked. Thanks!!

Problems:
- too casual (might not match your brand)
- overconfident (“definitely”) and vague (“not too bad”)
- no clear next step for approval

After (with SOUL.md boundaries + formatting)

Summary: Draft reply to lead (no price promises; scheduling-first).

Draft (ready to paste):
Hi — thanks for reaching out. We can likely fit you in this week. Pricing depends on the size/condition of the space, so I don’t want to guess.

If you share (1) your address/zip, (2) bedrooms/bathrooms (or approximate square footage), and (3) your preferred day/time window, I can confirm availability and a quote.

Question: Want me to send this?

That’s the goal: your agent produces drafts that are consistent, safe, and easy to approve.

---

Related Guides

• OpenClaw Workspace Files Explained: /guides/02-workspace-files
• Setting Up OpenClaw in a Weekend: /guides/05-setup-guide
• OpenClaw Multi-Agent Guide: /guides/07-multi-agent