Prompt Engineering
Cheat Sheet

Assign a persona or expertise to the model to shape tone, depth, and domain focus. The model adapts vocabulary, reasoning style, and assumptions to match.

When to use: Whenever you need domain-specific expertise, consistent tone, or a specific perspective. Works especially well for code review, writing, and analysis.

Provide 2–5 input→output examples before your actual request. The model learns the pattern and applies it consistently. More reliable than describing the format in words.

When to use: Classification, formatting, style matching, data transformation. Essential when zero-shot results are inconsistent.

Ask the model to reason step-by-step before giving a final answer. Dramatically improves accuracy on math, logic, and multi-step reasoning tasks.

When to use: Math problems, logical deductions, complex analysis, debugging. Add "Let's think step by step" as a simple trigger.

Break your request into numbered steps to control execution order and ensure nothing is skipped. Works like pseudocode for the model.

When to use: Multi-part tasks, workflows, any time order matters. Prevents the model from skipping steps or combining them incorrectly.

Provide rich context and constraints without examples. Works well when you clearly define what you want, including format, length, tone, and audience.

When to use: When the task is clear enough that examples aren't needed. Good for creative tasks, simple transformations, and when you have specific constraints.

Use clear delimiters (triple quotes, XML tags, markdown headers) to separate instructions from data. Prevents prompt injection and improves clarity.

When to use: Whenever your prompt includes user-supplied data, long documents, or multiple distinct sections. Critical for security.

The model explores multiple reasoning paths, evaluates each, and selects the best. Like CoT but branching — better for problems with multiple valid approaches.

When to use: Strategy decisions, architecture design, complex debugging where multiple hypotheses exist. Higher token cost but better outcomes.

Generate multiple independent answers, then pick the most common result. Reduces variance and catches errors through majority voting.

When to use: High-stakes calculations, factual questions, when you need confidence in the answer. Best with temperature > 0 for diversity.

Ask the model to generate an answer, critique it against specific criteria, then revise. Implements a feedback loop within a single prompt.

When to use: Quality-critical content, customer-facing text, anywhere you'd normally review output manually.

Use the model to generate or improve prompts. Feed it your goal and let it craft the optimal prompt — then use that prompt. Recursion for prompts.

When to use: When you're stuck crafting a prompt, building prompt templates for production, or optimizing existing prompts for better results.

Interleave reasoning and tool-use actions. The model thinks about what to do, executes an action, observes the result, then reasons about the next step.

When to use: Agent workflows, tool-augmented tasks, research questions requiring multiple lookups.

Break complex tasks into a pipeline of simpler prompts, where each step's output feeds the next. More controllable and debuggable than one mega-prompt.

When to use: Complex workflows, content pipelines, anywhere a single prompt would be unreliably long. Each step can use a different model or temperature.

Force the model to output valid JSON. Use the API's response_format parameter when available, or specify the schema in your prompt.

Pro tip: Provide the exact JSON schema with example values. Use response_format: {type: "json_object"} in the API for guaranteed valid JSON.

Use XML-style tags to structure both input and output. Especially effective with Claude, which was trained to respect XML tag boundaries.

Pro tip: Claude particularly excels with XML. Use for separating instructions from data and for multi-part outputs. Also helps prevent prompt injection.

Request output in markdown for human-readable structured content. Great for documentation, reports, and content that will be rendered in a UI.

Pro tip: Specify which markdown elements to use. Models default to markdown, but explicit instructions prevent inconsistent heading levels or missing tables.

Use the API's function calling / tool use feature to get perfectly typed responses. The model fills in a predefined schema — no parsing needed.

Pro tip: Most reliable way to get structured data. Use OpenAI's tools, Anthropic's tool_use, or Google's function_calling.

• ✓ System message is king — put persona & constraints there
• ✓ Excellent at response_format: json_object
• ✓ Function calling is best-in-class
• ✓ Follows numbered instructions very well
• ⚠ Can be verbose — add "be concise" explicitly
• ⚠ May refuse edge cases — rephrase as hypothetical
• ✓ o1/o3 models: skip CoT prompting — they do it internally

• ✓ XML tags are a superpower — use <tags> everywhere
• ✓ Excels at long documents (200K context)
• ✓ Best at following nuanced, detailed instructions
• ✓ Strong at code — give full file context
• ✓ Extended thinking for hard reasoning tasks
• ⚠ Can be overly cautious — frame requests clearly
• ✓ Prefill assistant message to steer output format

• ✓ Excellent multimodal — images, video, audio natively
• ✓ Massive context (1M+ tokens) — dump everything in
• ✓ Great at grounded search with Google integration
• ✓ Strong at structured data and tables
• ⚠ Less reliable at strict format compliance
• ⚠ May hallucinate on niche topics — provide sources
• ✓ Use system instructions for consistent persona

📝 Summarization

Always specify: length, audience, focus area, and what to omit.

🔍 Extraction

Define exact fields, types, and how to handle missing/ambiguous data.

🏷️ Classification

Enumerate all valid categories. Add confidence scores for filtering.

✨ Generation

The more constraints you give, the better the output. Underconstrained = generic.

📊 Analysis

Structure the output. Ask for prioritized findings and unknowns.

💻 Code

Specify language, I/O types, edge cases, and conventions. Always ask for tests.

Top-P (Nucleus Sampling)

Controls the cumulative probability pool. Lower = more focused. Usually set one of temp OR top-p, not both.

Max Tokens

Set output length limits. Prevents runaway generation and controls cost. 1 token ≈ ¾ of a word in English.

Other Parameters

🎯 Prompt Compression

• → Remove filler words: "I would like you to please" → just state the task
• → Use abbreviations in system prompts (models understand shorthand)
• → Replace verbose instructions with 1 example (few-shot > explaining)
• → Use structured data (JSON/YAML) instead of natural language for inputs

⚡ Context Management

• → Summarize conversation history instead of keeping full messages
• → Use RAG to inject only relevant chunks, not entire documents
• → Trim system prompts — most can be cut 50% without quality loss
• → Cache frequent system prompts (OpenAI prompt caching saves 50%)

🔀 Model Routing

• → Use smaller models for simple tasks (classification, extraction)
• → Route to larger models only for reasoning/creative tasks
• → GPT-4o-mini, Claude Haiku, Gemini Flash — 90% of tasks, 10% of cost
• → Build a classifier that picks the model per request

📊 Cost Math Cheat Sheet