Module 4: Writing Good Prompts

Why this matters

The quality of what your agent does depends heavily on how you ask it. A vague prompt gives a vague answer. A clear, structured prompt gives a useful one.

This module teaches you the patterns that work best with OpenClaw.

The basics: Be specific

Compare these two prompts:

Vague:

Summarize my emails

Specific:

Check my inbox for unread emails from the last 24 hours. Summarize each one in a single sentence. Group them by sender. Flag anything that looks urgent.

The specific version tells the agent:

• What to look at (unread, last 24 hours)
• How to format the output (one sentence per email)
• How to organize it (by sender)
• What to prioritize (urgent items)

Structure your prompts

A good prompt for OpenClaw has three parts:

1. Context — What situation is the agent working in?
2. Task — What exactly should it do?
3. Format — How should the output look?

Context: I have a meeting with the engineering team in 30 minutes.
Task: Check my calendar for today and summarize what I need to prepare.
Format: Bullet points, no more than 5 items.

Use SOUL.md for persistent instructions

Instead of repeating the same instructions every time, put them in your SOUL.md file. This file shapes how your agent behaves by default.

openclaw config edit soul

Example additions:

## Response style
- Keep responses concise — 3 sentences maximum unless I ask for detail
- Use bullet points for lists
- Always include the source when citing information
- Never use emojis unless I use them first

These instructions apply to every conversation automatically.

Templates for repeatable tasks

If you run the same type of prompt regularly, create a skill or save it as a template in your SOUL.md:

## Daily briefing template
When I ask for my daily briefing, always include:
1. Weather for my location
2. Calendar events for today
3. Unread messages summary
4. Top 3 news items in my areas of interest

Then you just say:

Give me my daily briefing

And the agent knows exactly what to do.

Chain-of-thought prompting

For complex tasks, ask the agent to think step by step:

I need to plan a product launch. Think through this step by step:
1. What are the key milestones?
2. What teams need to be involved?
3. What is a reasonable timeline?
Then give me a one-page summary.

This produces much better results than just asking "Help me plan a product launch."

Debugging bad outputs

If your agent gives poor results:

1. Too vague? Add more context and constraints
2. Too long? Specify a length limit ("in 3 sentences")
3. Wrong format? Show an example of what you want
4. Inconsistent? Move instructions to SOUL.md so they persist
5. Hallucinating? Ask it to cite sources or only use information from your tools

Hands-on: Improve a prompt

Try this exercise. Send this prompt to your agent:

openclaw agent --message "Tell me about the weather"

Now try the improved version:

openclaw agent --message "What is the current weather and forecast for the next 3 days in my location? Give me temperature, rain probability, and a one-sentence summary for each day."

Compare the results. The second version should be significantly more useful.