OpenClaw Prompting Guide

In ChatGPT, you write a prompt, get a response, and start over. Every conversation is a blank slate.

OpenClaw works differently. Your agent has persistent identity files that load every session. Instructions compound over time instead of resetting. Your agent remembers who it is, who you are, and what happened yesterday.

This means prompting OpenClaw is less about writing clever one-shot prompts and more about configuring a persistent system that gets better with every interaction. If you want to build a personal AI, understanding this difference is everything.

This is your agent's identity. Personality, tone, values, boundaries. Loaded at the start of every session.

Start simple. You can always add more personality as you figure out what you like.

The operating manual. Rules for memory, safety, file handling, group chat behavior, and external actions.

Your name, timezone, preferences, and context. So your agent knows you without you repeating yourself.

Long-term memory that persists across sessions. The agent reads it on startup and updates it with important information.

You do not need to write this manually. Your agent builds it over time as you interact.

Say "remember this" and OpenClaw writes it to memory. Next session, it already knows. "Remember: I prefer TypeScript over JavaScript" or "Remember: our API rate limit is 100 req/min."

OpenClaw is an agent, not a chatbot. Give it multi-step tasks: "Research the top 5 competitors for our product, summarize their pricing, and save a comparison table to a file." It will execute all steps.

Negative instructions are powerful. "Don't use marketing language." "Don't explain what you're doing, just do it." "Don't ask for confirmation on file reads." Put recurring ones in SOUL.md so they apply every session.

For recurring tasks, create custom skills. A skill is a SKILL.md file with task-specific instructions that the agent loads when relevant.

Browse pre-built skills at Best OpenClaw Skills.

• Being too vague - "help me with marketing" gives generic results. "Write 3 Twitter posts about our new feature launch, casual tone, under 280 chars" gets exactly what you need.
• Overloading SOUL.md - start with 10-15 lines. A 500-line SOUL.md wastes tokens and confuses the model. Add rules only when you notice problems.
• Not using memory - if you keep repeating the same context, tell the agent to remember it. That is literally what MEMORY.md is for.
• Treating it like a search engine - do not ask "what is X?" when you want action. Ask "research X and save a summary to a file."
• Ignoring USER.md - fill in your timezone and name at minimum. It makes every response more relevant.
• Not iterating - if the first response is not right, refine. "Make it shorter." "More casual." "Add numbers." The agent adjusts instantly.

The most powerful prompting in OpenClaw is not interactive - it is automated. Set up cron jobs with prompts that run on a schedule:

• Morning brief: "Summarize my unread emails, today's calendar, and any crypto price alerts."
• End of day: "Write today's memory file with what we worked on and any open tasks."
• Weekly review: "Review this week's memory files and update MEMORY.md with anything worth keeping long-term."

These prompts run without you. Your agent works in the background, maintaining context and delivering updates to your phone via Telegram, WhatsApp, or Discord.

• Edit SOUL.md - add 3-5 lines about the personality you want
• Edit USER.md - add your name and timezone
• Start chatting - the agent reads your files automatically
• Say "remember this" - when you share something important
• Iterate - update SOUL.md when something annoys you

For the full setup, see the beginner guide. For advanced configuration, check the config reference. To pick the right model, see the AI model guide. To switch models, see how to change models.