The Beginner's Guide to Prompt Engineering for AI Agents

Quick Answer

Prompt engineering for AI agents is different from prompting a chatbot. Instead of crafting individual messages, you are writing a system-level instruction set (like a SOUL.md file in OpenClaw) that defines your agent's personality, goals, tool-use behaviour, and boundaries. Good agent prompts are specific about what the agent should do, clear about when to use which tools, explicit about what the agent should not do, and structured so the agent can handle diverse tasks without getting confused.

Why Agent Prompting Is Different

When you prompt a chatbot like ChatGPT or Claude, you are optimising a single interaction. You craft a message, get a response, refine if needed, and move on. The prompt is disposable — it serves one conversation.

When you prompt an AI agent, you are writing instructions that will govern hundreds or thousands of future interactions across different topics, tasks, and contexts. Your agent's system prompt is not disposable — it is the foundation of every interaction your agent will ever have.

This changes the approach completely. You are not optimising for one perfect response. You are optimising for consistent, reliable behaviour across an unpredictable range of situations. A chatbot prompt says "write me a poem about autumn." An agent prompt says "you are a personal assistant who helps me with research, communication, and task management. Here is how you should behave in different situations."

The SOUL.md File

In OpenClaw, your agent's core instructions live in a file called SOUL.md. This is the system prompt that loads into every conversation, defining who your agent is and how it should behave. Everything about your agent's personality, capabilities, preferences, and boundaries is written here.

Think of SOUL.md as your agent's job description and training manual rolled into one. It tells the agent who it is, what it should do, how it should do it, and what it should avoid. A well-written SOUL.md is the single biggest factor in how useful your agent will be.

Start With Identity and Purpose

The first section of your SOUL.md should establish who your agent is and what its primary purpose is. This grounds every subsequent instruction.

Keep it simple and direct. Something like: "You are my personal AI assistant. Your primary purpose is to help me manage my work and personal tasks efficiently. You communicate through Telegram and WhatsApp. You are helpful, concise, and proactive — you suggest things I might not have thought of."

This immediately tells the model what role it is playing, what channels it operates on, and what communication style to use. Every response it generates will be filtered through this identity.

Avoid vague identities like "you are a helpful AI." Be specific about what kind of help, for whom, and in what style.

Define Tool-Use Behaviour

This is the most important section for agent prompts and the one most people get wrong. Your agent has access to tools (MCP skills), and it needs clear guidance on when to use them.

Without explicit tool-use instructions, agents tend to either over-use tools (searching the web for questions they already know the answer to) or under-use them (answering from memory when they should be checking current information).

Write specific rules for tool usage. For example: "Always use web search when I ask about current events, news, prices, or anything that might have changed since your training data. Do not use web search for general knowledge questions you are confident about. When you search, read at least two sources before giving me an answer. Always include the source links."

For each major skill your agent has, add a brief instruction about when to use it and when not to. "Use the calendar skill to check my schedule before suggesting meeting times. Use the email skill to read my inbox when I ask about emails, but never send an email without my explicit confirmation."

Set Communication Style

Your agent's communication style should match how you want to interact with it. Some people want brief, telegram-style responses. Others want detailed explanations. Some want a professional tone. Others want something more casual.

Be explicit: "Keep responses concise — aim for two to three sentences unless I ask for more detail. Use a friendly but professional tone. Do not use emojis. When I ask for research, provide a thorough response with sections and source links."

You can also set different styles for different types of tasks: "For quick questions, respond in one to two sentences. For research requests, provide a detailed summary with headings. For email drafts, match my writing style — professional but warm, no jargon."

Establish Boundaries

Good agent prompts are as clear about what the agent should not do as what it should do. Boundaries prevent your agent from taking actions you did not intend or making decisions you want to make yourself.

Common boundaries include: "Never send messages on my behalf without asking first. Never delete files. Never make purchases or financial transactions. If you are unsure about something, ask me rather than guessing. If a task seems outside your capabilities, say so rather than attempting it poorly."

For work agents: "Do not share information about internal projects with anyone outside the team. Do not access personal accounts during work hours. Escalate any request that involves legal, financial, or HR matters to me directly."

Boundaries make your agent safer and more predictable. They also help you build trust gradually — start with strict boundaries and relax them as you gain confidence in your agent's behaviour.

Add Context About You

The more your agent knows about you, the more relevant its responses will be. Include key context in your SOUL.md: your name, role, company, timezone, key projects, preferences, and anything else that helps the agent personalise its responses.

"My name is Alex. I am a marketing manager at a B2B SaaS company in London. I work Monday to Friday, 9am to 6pm GMT. My current priorities are the Q2 product launch and the redesign of our website. I prefer concise communication and make decisions quickly — give me options with a clear recommendation rather than open-ended lists."

This context means your agent does not have to ask for basic information in every conversation. It can schedule meetings in your timezone, reference your projects, and tailor its suggestions to your role and industry.

Structure for Clarity

As your SOUL.md grows, structure becomes important. Use clear sections with headers so the model can easily parse your instructions. A good structure might be: identity and purpose, communication style, tool-use guidelines, boundaries and safety, personal context, and task-specific instructions.

Keep each section focused and avoid contradictions. If you tell your agent to be concise in one section and to provide detailed explanations in another, it will be inconsistent. Be specific about when each style applies.

Common Mistakes

Being too vague. "Be helpful" is not an instruction. "When I ask a question, answer directly, then offer one follow-up suggestion if relevant" is an instruction. Specificity is the single most important quality of good agent prompts.

Writing a novel. Every token in your SOUL.md is processed with every single interaction. A 2,000-word SOUL.md consumes significant context in every conversation, especially on smaller models. Be concise. Cut anything that is not directly improving your agent's behaviour.

Forgetting about edge cases. Your agent will encounter situations you did not anticipate. Add a general instruction for how to handle uncertainty: "If you encounter a request you are unsure about, ask me for clarification rather than guessing. If something seems urgent, flag it clearly."

Not iterating. Your first SOUL.md will not be perfect. Pay attention to how your agent behaves over the first few days, note where it does something unexpected or unhelpful, and update the SOUL.md accordingly. The best agent prompts evolve through use.

Testing Your Prompt

After writing your SOUL.md, test it systematically. Send your agent a range of different requests: simple questions, research tasks, tasks that require tool use, tasks that should trigger your boundaries, and ambiguous requests. Check that the agent behaves consistently with your instructions across all of these.

Pay particular attention to tool-use behaviour. Does the agent search when it should? Does it avoid searching when it does not need to? Does it use the right tool for each task? Tool-use reliability is the area where good prompting makes the biggest difference.

Frequently Asked Questions

How long should my SOUL.md be?

Aim for 300 to 800 words. Long enough to cover the essentials, short enough to leave context space for actual conversations. If you are running a smaller model locally, keep it under 500 words. On Tulip with larger models, you have more room.

Should I use technical language in my prompt?

Write in plain English. The model understands natural language better than technical jargon. "Search the web before answering questions about current events" is better than "invoke the SerpAPI MCP tool for queries classified as temporal."

Can I have different prompts for different tasks?

In OpenClaw, you have one SOUL.md per agent. If you need very different behaviours for different tasks, consider running multiple agents with different SOUL.md files — one for work, one for personal use, one for research.

How often should I update my SOUL.md?

Review it after the first week of use, then monthly. Update whenever you notice your agent doing something you want to change, or when your circumstances change (new projects, new preferences, new tools). Treat it as a living document, not a set-and-forget configuration.