Download the full guide.

The foundation of every good prompt. Context + Task + Format = Quality Output. Get this right before reaching for anything else.

• Always provide relevant context first
• Be specific about the task — vague in, vague out
• Define output format explicitly

Provide 2–5 input/output examples to pattern-match the model's behavior. Faster than writing detailed instructions for consistent formatting.

• Use 2–5 examples for best results
• Keep examples diverse but consistent in structure
• Include edge cases if your task has them

Force the model to show its reasoning before answering. Dramatically improves accuracy on multi-step logic, math, and analytical tasks.

• "Let's think step by step" is a reliable trigger
• Ask for intermediate reasoning, not just the answer
• Verify each reasoning step in the output

Assign a specific expert persona to shape tone, vocabulary, and depth. "Act as a senior X with Y years of Z" outperforms generic requests.

• Be specific about expertise level and years
• State the audience — expert role + lay audience = clear explanations
• Add behavioral constraints ("be concise", "no jargon")

Reasoning + Acting. Interleave thought and action for tool-augmented workflows. The backbone of agentic AI systems.

• Use for tool-augmented / function-calling workflows
• Structure strictly as Thought → Action → Observation
• Useful for API integration and multi-step retrieval

Force specific output formats for reliable downstream parsing. Provide the exact schema — don't describe it, show it.

• Provide the exact schema, not a description of it
• Use "No prose, only JSON" for strict formatting
• Validate output structure programmatically

Define what NOT to do as clearly as what TO do. Negative constraints narrow the output space and prevent common failure modes.

• Use negative constraints sparingly but precisely
• Set word/character limits to prevent over-generation
• Define what counts as out-of-scope

Generate multiple independent answers and take the majority vote. Reduces variance on critical decisions where a single run may be unreliable.

• Use higher temperature (0.7–1.0) to ensure diversity between runs
• Ask the model to evaluate its own options
• Best justified on high-stakes, low-frequency decisions

Control randomness and creativity through sampling parameters. The right temperature is determined by the task, not the vibe.

• Temp 0 for code/facts; 0.7–0.9 for creative writing
• Adjust Top P instead of Temperature for finer control
• Set max tokens to prevent runaway generations

Use clear delimiters to separate instructions from content. Prevents prompt injection and makes long, multi-part inputs unambiguous.

• Triple quotes for long text blocks
• XML tags for complex multi-part inputs
• Markdown code blocks for code snippets

Explore multiple reasoning paths simultaneously and backtrack when dead ends are reached. Simulates deliberate planning rather than linear generation.

• Best for complex planning and creative problem-solving
• Simulates multi-agent deliberation in a single prompt
• Allows exploration, evaluation, and backtracking

Set global behavior rules that persist across the entire conversation. The right place to define persona, safety boundaries, and response defaults.

• Use for persistent behavioral instructions
• Define scope, persona, and safety limits here
• Keep system prompts focused — longer ≠ better