AI agent prompting strategies that actually improve productivity

The difference between a tool and an assistant

Most people interact with AI the same way they use a search engine: type a question, read the answer, move on. This works but leaves significant value on the table. The difference between an AI that saves you five minutes a day and one that saves an hour comes down almost entirely to how you structure your requests and how well the agent knows your context.

This guide covers techniques that work specifically for personal agents like OpenClaw — systems that run continuously and accumulate context about you over time.

The system prompt: your most important configuration

The system prompt is the persistent instruction that prefixes every conversation with your agent. It's where you define who your agent is and how it should behave. Most people either leave it at the default or write a generic instruction. Both are missed opportunities.

An effective system prompt for a personal agent includes:

• Your professional context: What you do, what tools you use, what projects you're working on
• Your communication preferences: Concise or detailed responses, bullet points or prose, technical depth level
• Standing instructions: "Always ask before making assumptions about deadlines" or "When I ask for feedback, be direct and critical"
• Important context: Time zone, language preferences, recurring priorities

Example system prompt excerpt:

You are my personal AI assistant. I'm a freelance software developer based in Madrid, Spain (UTC+1). I primarily work in TypeScript and Python. I value directness over diplomacy — when I ask for feedback, I want honest critique, not reassurance. Keep responses concise unless I ask for depth. When I share URLs, summarize the key points proactively without being asked.

Task decomposition for complex requests

AI models handle complex requests better when broken into explicit steps. Instead of:

Help me plan my product launch.

Try:

I'm planning a product launch for [X]. Help me structure this in three parts:
1. A timeline for the 4 weeks before launch
2. A checklist of deliverables by category (marketing, technical, legal)
3. A list of risks I might be overlooking

The second version consistently produces more useful output because it constrains the scope of each section and forces the model to organize its response logically.

Using memory effectively

One of the biggest advantages of a personal agent over ChatGPT is persistent memory. But memory is only useful if you actively build it. When you discover something your agent should always know, tell it explicitly:

Remember: I prefer Anthropic models for writing tasks and DeepSeek for code review. Use this when I don't specify a model.

Remember: The client at Acme Corp is called Sandra, and she prefers communication by email, not messaging.

Then periodically ask your agent to show you what it remembers (/memory command) and correct anything outdated. Treat agent memory like you would onboarding a human assistant: invest time once, benefit continuously.

The role-assignment technique

Asking the model to adopt a specific role often produces better-calibrated responses than general requests:

• "As a skeptical investor, poke holes in this business idea..."
• "As a senior engineer reviewing this code, what would you flag..."
• "As someone unfamiliar with my industry, explain why this document is confusing..."

This works because roles carry implicit context about what kind of response is appropriate. A "skeptical investor" role signals that you want challenges and risks, not validation.

Iterative refinement vs. single-shot requests

For important outputs — a document, a plan, a communication — treat the first draft as a starting point, not an end product. A productive iteration pattern:

1. Generate the initial draft with a clear request
2. Ask for specific changes: "Make the third paragraph more concrete, replace the abstract claims with examples"
3. Ask for a different perspective: "Now argue the opposite position"
4. Ask for compression: "Now cut this by 30% without losing the key points"

Each iteration improves the output in a targeted way. This is faster than rewriting from scratch and produces better results than trying to specify everything in the initial prompt.

Calibrating response length

AI models tend toward verbosity. Get in the habit of specifying the format you want:

• "In one sentence:"
• "Give me three bullet points, no explanation:"
• "Explain this as if I have 30 seconds to understand it:"
• "Full analysis, I have time to read thoroughly:"

Adding format instructions to your system prompt as defaults saves time on every routine interaction.

Building a personal prompt library

Once you find prompting patterns that work well for recurring tasks, save them. OpenClaw supports custom command shortcuts — you can define /weekly-review as a full prompt that triggers your end-of-week analysis, or /email-draft as a template for drafting professional emails in your style.

Think of these as macros for your agent. The upfront investment in writing a good prompt once pays back every time you use the shortcut.