DEV Community

Midas Tools
Midas Tools

Posted on

How to Write a SOUL.md That Makes Your AI Agent Actually Useful

#webdev #ai #productivity #tutorial

If you are building AI agents — whether with Claude, GPT, LangChain, or any framework — the most important file is not your code. It is your agent's identity document.

I call it SOUL.md. It is the system prompt that defines who your agent is, what it can do, and what it must never do. A vague SOUL.md produces a vague agent. A specific one produces an agent that actually gets work done.

After studying agents that generate real revenue (including autonomous agents running entire businesses), here is what separates agents that work from agents that do not.

The 5 Components of a Good Agent Identity

1. Specific Identity (not "helpful assistant")

Bad:

You are a helpful AI assistant.
Enter fullscreen mode Exit fullscreen mode

Good:

You are Atlas, the autonomous operations agent for Sunrise Coffee.
Company: Shopify store, 500 orders/month, 3-person team.
Your role: reduce founder's daily admin from 4 hours to 30 minutes.
Enter fullscreen mode Exit fullscreen mode

The difference: a "helpful assistant" has no context and no goals. Atlas knows the business, the scale, and the mission.

2. Defined Skills (not vague capabilities)

Bad:

You can do anything the user asks.
Enter fullscreen mode Exit fullscreen mode

Good:

Skills:
- Product listing optimization (SEO titles, descriptions, tags)
- Customer support automation (response within 5 minutes)
- Social media posting (3x daily at 9am/12pm/6pm)
- Competitor price tracking (weekly report every Monday)
- Inventory alerts (notify when stock < 10 units)
Enter fullscreen mode Exit fullscreen mode

List every skill. If it is not listed, the agent should not attempt it. This prevents hallucinated capabilities and keeps the agent focused.

3. Clear Constraints (the most important section)

Without constraints, your agent WILL do something you do not want. It is not a question of if, but when.

Constraints:
- Never discount more than 15% without human approval
- Never spend more than $50/day on any paid action
- Never delete customer data
- Always escalate negative reviews immediately
- Log every customer interaction for audit trail
- Never make claims about product safety or health benefits
Enter fullscreen mode Exit fullscreen mode

The more specific your constraints, the safer your agent runs autonomously.

4. Operating Schedule

Agents without schedules either run constantly (wasting resources) or inconsistently (missing opportunities).

Schedule:
- 09:00 — Check overnight orders, respond to support tickets
- 12:00 — Post to social media, check inventory levels
- 15:00 — Competitor price check, update listings if needed
- 18:00 — Send daily summary to founder
- Continuous — Monitor for negative reviews (respond within 5 min)
Enter fullscreen mode Exit fullscreen mode

5. Escalation Rules

Your agent needs to know when to stop and ask a human.

Escalation:
- Customer requests refund > $100 → notify founder
- Any legal or compliance question → stop, do not respond
- System error or API failure → log and alert immediately
- Any request outside defined skills → politely decline
Enter fullscreen mode Exit fullscreen mode

Industry-Specific Templates

Different businesses need different agent configurations. Here are starting points:

E-commerce: Focus on listing optimization, customer support, inventory alerts, review management. Key tools: Shopify API, Stripe, social media APIs.

SaaS: Focus on onboarding sequences, churn detection, support triage, usage analytics. Key tools: Stripe, Intercom, PostHog, email.

Content Creator: Focus on content calendar, trend monitoring, engagement tracking, sponsorship management. Key tools: YouTube API, social schedulers, Google Trends.

Freelancer: Focus on proposal generation, client updates, time tracking, invoicing. Key tools: email, Notion, calendar, Stripe.

Real Estate: Focus on lead follow-up, listing management, market analysis, showing scheduling. Key tools: MLS, CRM, email, calendar.

Generate Your Config in 60 Seconds

I built a free SOUL.md Generator that handles all of this. Pick an industry preset or go custom, and it generates a production-ready identity document with:

  • Agent name and personality
  • Industry-specific skills
  • Tool integrations
  • Operating schedule
  • Constraints and escalation rules
  • Completeness tracking (so you know what is missing)

No signup required. Generate, copy, paste into your agent framework.

The Checklist

Before deploying any agent, make sure your SOUL.md has:

  • [ ] Specific identity (name, company, role)
  • [ ] Measurable mission (not "be helpful")
  • [ ] Defined skills (explicit list, not "anything")
  • [ ] Clear constraints (spending limits, content rules, escalation triggers)
  • [ ] Operating schedule (when to run, when to rest)
  • [ ] Escalation rules (when to stop and ask a human)
  • [ ] Tool integrations (which APIs, which permissions)
  • [ ] Logging requirements (what to track, where to store it)

If your agent is missing any of these, it will eventually fail in a way you did not anticipate.

Generate your SOUL.md free | All AI tools at midastools.co/tools

Top comments (0)