How many few-shot examples should I include?

Two to four examples work well for most tasks. More examples improve consistency but consume more tokens. If your agent handles diverse query types, use two examples per type rather than eight examples for a single type.

Does chain-of-thought increase response time?

Yes, because the model generates more tokens when reasoning step by step. If speed is critical, you can instruct the agent to reason internally and only output the final answer. This preserves accuracy while reducing visible response length.

Can I use different prompts for different channels?

Yes. You can configure channel-specific system prompts. For example, a more concise prompt for SMS where message length matters, and a detailed prompt for web chat where users expect thorough answers.

Advanced

Prompt Engineering: Advanced Techniques

Refine your OpenClaw agent's system prompts with advanced techniques that improve accuracy, reduce hallucinations, and produce consistently structured outputs.

Deploy OpenClaw See How It Works

What You Will Get

By the end of this guide, you will have a toolkit of advanced prompt engineering techniques that make your OpenClaw agent more reliable, accurate, and consistent. You will know when to use chain-of-thought reasoning, how to craft effective few-shot examples, and how to enforce structured output formats.

Prompt engineering goes far beyond writing a simple instruction. The way you structure a system prompt directly affects how the model reasons, what it includes in its answer, and how it handles edge cases. Small changes in phrasing can have a dramatic impact on output quality.

You will apply each technique to your agent's system prompt, test the results, and iterate until the output meets your standards. The result is an agent that handles complex queries with precision and delivers answers in exactly the format you need.

Step-by-Step Setup

Follow these steps to apply advanced prompting techniques.

Audit Your Current Prompts

Start by reviewing your agent's existing system prompt and any task-specific prompts. Identify areas where the agent produces inconsistent, verbose, or inaccurate responses. These problem spots are where advanced techniques will have the most impact.

Add Chain-of-Thought Instructions

Instruct the agent to reason step by step before giving a final answer. Add a line like 'Think through the problem step by step before responding.' This technique is especially effective for math, logic, and multi-factor analysis tasks. It reduces errors by forcing the model to show its work.

Include Few-Shot Examples

Add two to four example input-output pairs directly in the system prompt. These examples teach the agent the exact format, tone, and level of detail you expect. Choose examples that cover common cases and at least one edge case. The agent will pattern-match against these examples for future queries.

Enforce Structured Output

If you need JSON, Markdown tables, or bullet-point lists, specify the exact output format in the system prompt. Include a template with placeholder values so the agent knows precisely what to produce. Add an instruction like 'Always respond in the following JSON format' followed by the template.

Define Boundaries and Constraints

Explicitly state what the agent should not do. For example, 'Do not make up information. If you are unsure, say so.' or 'Never include personal opinions.' Clear boundaries prevent the agent from drifting off-topic or producing unreliable content.

Use Role and Persona Framing

Assign the agent a specific role at the start of the system prompt. For example, 'You are a senior data analyst who explains findings clearly to non-technical stakeholders.' This framing shapes the agent's tone, vocabulary, and level of detail throughout the conversation.

Test and Iterate

After applying each technique, test with a set of representative queries. Compare the new outputs against the previous ones. Keep a changelog of prompt modifications and their effects so you can revert changes that do not improve quality.

Tips and Best Practices

One Change at a Time

Modify one aspect of the prompt per iteration. If you change the persona, add examples, and restructure the format all at once, you cannot tell which change improved or worsened the output.

Use Delimiter Tokens

Separate different sections of your prompt with clear delimiters like triple dashes or XML-style tags. This helps the model distinguish between instructions, examples, and user input, reducing confusion.

Keep Prompts Under 1,500 Tokens

Longer system prompts are not always better. Concise prompts leave more room for conversation context and retrieved knowledge. Remove redundant instructions and consolidate overlapping rules.

Version Control Your Prompts

Store prompt versions with timestamps and performance notes. When an update degrades quality, you can quickly roll back to the previous version. This is especially important for production agents serving real users.

Frequently Asked Questions

Agent Personality Tuning Model Response Quality Custom Tool Creation

Ready to get started?

Deploy your own OpenClaw instance in under 60 seconds. No VPS, no Docker, no SSH. Just your personal AI assistant, ready to work.

Deploy OpenClaw View Pricing

Starting at $24.50/mo. Everything included. 3-day money-back guarantee.