Your agent isn't broken. Your SOUL.md is.
I've deployed dozens of AI agents. WhatsApp bots, Telegram assistants, Discord helpers — you name it. And for months, I kept hitting the same wall:
The agent would work perfectly for 3-4 messages. Then it would completely lose character.
Sound familiar?
Maybe your agent starts responding generically. Maybe it ignores the constraints you carefully defined. Maybe it suddenly performs actions you never configured it to do.
I tried everything. Longer system prompts. Shorter system prompts. JSON schemas. XML tags. More examples. Fewer examples.
Nothing stuck.
Then I created the LEONIDAS framework, and my agents finally work.
The Prompt Problem Nobody Talks About
Here's what most agent tutorials don't tell you: the structure of your prompt matters more than the content.
Most developers write SOUL.md files like this:
You are a sales assistant. Help users with their questions
and try to sell them our products.
That's not a prompt. That's a prayer.
No constraints. No business logic. No behavior rules. No fallback handling. Your LLM has zero guidance on edge cases — and in production, everything is an edge case.
The LEONIDAS framework solves this with 8 distinct pillars:
| Pillar | Purpose |
|---|---|
| L – Persona | Who is the agent? Background, expertise, track record |
| E – Objective | What's the one mission? Every message should advance it |
| O – Tone & Format | Channel-specific formatting (WhatsApp ≠ Telegram ≠ Discord) |
| N – Constraints | What should the agent never do? |
| I – Business Logic | Decision trees, qualification criteria, routing rules |
| D – Structure | Conversation flow (Greeting → Discovery → Action → Close) |
| A – Human Behavior | Psychology and relationship dynamics |
| S – Multipurpose | Platform adaptations |
This isn't theory. It's battle-tested across thousands of OpenClaw deployments.
Before & After: The Difference Is Night and Day
Let me show you a real example — a WhatsApp sales agent.
❌ Before (Generic Prompt)
You are a sales assistant. Help users with their questions
and try to sell them our products.
Result? Generic responses. No qualification. No structure. No close rate.
✅ After (LEONIDAS SOUL.md)
[L – PERSONA]
You are Sofia, a senior sales consultant with 8 years of
experience closing high-ticket B2B deals over WhatsApp.
You have a 34% close rate on qualified leads and specialize
in consultative selling — never pushy, always value-first.
[E – OBJECTIVE]
Your sole mission is to qualify inbound leads, understand
their pain points, and guide them toward booking a discovery
call. Every message must move the conversation one step
closer to a booked meeting.
[O – TONE & FORMAT]
Conversational, warm, and direct. WhatsApp messages must be
under 3 sentences unless the user asks for detail. Use line
breaks, not paragraphs. Never use corporate jargon.
[N – CONSTRAINTS]
Do not discuss pricing before qualification. Do not send more
than 2 messages without a response. If a lead goes cold after
3 follow-ups, pause and flag for human review.
[I – BUSINESS LOGIC]
Qualification criteria: budget >$5K/month, decision-maker,
timeline <90 days. If all three are met, offer a 20-minute
discovery call. If not, nurture with value content.
[D – STRUCTURE]
Every sales conversation follows: Greeting → Pain Discovery →
Value Statement → Qualification → CTA (call booking).
[A – HUMAN BEHAVIOR]
People buy from people they trust. Build rapport before
pitching. Mirror the user's communication style — if they're
brief, be brief. If they're detailed, match their depth.
[S – MULTIPURPOSE]
This SOUL.md is optimized for WhatsApp. For Telegram, increase
message length by 20%. For web chat, enable markdown formatting.
Same agent. Same model. Completely different results.
The LEONIDAS version stays in character. It follows the conversation structure. It knows when to escalate. It adapts to the channel.
Generate Your SOUL.md in 60 Seconds
The best part? You don't have to write this yourself.
ASK LEONIDAS generates production-ready SOUL.md prompts automatically:
- Describe the problem — What's your agent doing wrong?
- Set the context — Role, industry, desired outcome
- Get your SOUL.md — Complete 8-pillar LEONIDAS prompt
No signup required. No account needed. Just paste and deploy.
It also supports:
- AGENTS.md for multi-agent coordination
- Skill prompts for specific capabilities (inbox management, lead qualification, appointment booking)
- Workflow prompts for sequential step logic
Why This Matters for the Agentic Economy
We're entering an era where AI agents don't just answer questions — they act. They book appointments. They qualify leads. They manage tickets. They coordinate with other agents.
This is the agentic economy, and it demands a new level of trust infrastructure. Your agents need:
- Clear identity — Who is this agent, really?
- Defined capabilities — What can it do? What can't it?
- Verifiable behavior — Does it follow the rules?
- Graceful escalation — Does it know when to stop?
The LEONIDAS framework bakes all of this into the prompt itself. No external guardrails needed. No post-hoc filtering. The agent knows its boundaries because they're encoded in its DNA.
If you're building production agents, you should also be following developments in agent trust and verification — because the standards are evolving fast.
TL;DR
- Your OpenClaw agent isn't broken — your SOUL.md is
- Generic prompts produce generic (and inconsistent) behavior
- The LEONIDAS framework uses 8 pillars to create bulletproof agent prompts
- Generate your SOUL.md in 60 seconds — free, no signup
- As we build the agentic economy, prompt architecture becomes infrastructure
Try It Now
👉 askleonidas.com/wizard?template=openclaw-soul
Free. No signup. 60 seconds.
If you're deploying OpenClaw agents, this will save you hours of prompt iteration — and your agents will finally stay in character.
What prompt frameworks are you using for your agents? Drop a comment below — I'm always looking for new approaches to test.
Tags: #ai #agents #openai #llm #promptengineering #chatbots #automation
Top comments (0)