AI Agent System Prompts

Assign a specific role or persona that focuses the AI's behavior and sets appropriate tone. Even a single sentence makes a significant difference.

Clearly define what the AI should accomplish, including the goal, constraints, and success criteria. Explicit is better than implicit.

Provide detailed rules for interaction, output format, and behavior constraints. Use structured formats (XML tags, numbered lists) for clarity.

Include 3-5 diverse, relevant examples showing exactly what you want. More examples improve performance on complex tasks.

Specify exact output structure, format, and style. Be specific about tone, structure, and any special formatting requirements.

If using tool/function calling, provide clear descriptions of available tools, when to use them, and expected parameters.

❌ Avoid: "Be helpful" or "Do the right thing". ✅ Instead: Use specific, measurable instructions with examples.

❌ Don't include conflicting constraints. ✅ Prioritize and clearly rank rules if they might conflict.

❌ Don't assume the AI knows your domain. ✅ Provide necessary background and context explicitly.

❌ Avoid prose-heavy, hard-to-parse instructions. ✅ Use structured formats (XML, markdown, clear sections).

❌ Don't skip examples. ✅ Always provide concrete examples of desired output.

❌ Don't oversimplify complex tasks. ✅ Break tasks into steps with clear acceptance criteria.

Reasoning + Acting: Agent generates verbal reasoning (Thought), takes actions (Action), observes results (Observation), then iterates. Perfect for complex reasoning with external tools. Works well with search, math, and multi-step problems.

Step-by-Step Reasoning: Agent breaks problems into intermediate steps showing work. Zero-shot variant just adds "think step by step". Few-shot includes examples. Best for logic, math, and reasoning tasks.

Tool Integration: AI decides which tools/functions to call and with what parameters. System prompt defines available tools, parameters, and decision-making logic. Essential for agents that interact with APIs and systems.

Specialized Roles: Different agents with specific roles collaborate (e.g., researcher + writer + editor). Each has a specialized prompt. Uses structured communication through documents or messages.

Retrieval Augmented Generation: Agent retrieves relevant documents/context before answering. System prompt guides what to search for and how to synthesize results. Critical for knowledge-grounded responses.

Learning by Example: Include 3-5 input/output examples in the prompt. Model learns the pattern and applies to new inputs. More examples = better performance on complex tasks.