Prompt Engineering Guide: 12 Techniques That Actually Work

Prompt engineering is the skill of writing instructions that get reliable, useful output from AI models like ChatGPT, Claude and Gemini. The techniques that actually work are simple: be specific about the goal, give context, show examples, structure the prompt with clear sections, break big tasks into steps, and test prompts against clear success criteria. Both OpenAI and Anthropic recommend these same core practices in their official guides.

Below are 12 techniques, each with a short explanation, a before and after example, and when to use it. They work in chat apps and in the API, and several of them also lower your token bill.

What prompt engineering is (and is not)

A prompt is everything you send the model: instructions, context, examples and the question. Prompt engineering means designing that input deliberately so the output is accurate, consistent and in the format you need. It is not about magic phrases or secret tricks. Modern models are good at following plain instructions; most of the gains come from removing ambiguity.

Anthropic’s documentation makes a useful point: “Not every success criteria or failing eval is best solved by prompt engineering.” Sometimes you can improve speed and cost more easily by picking a different model. Keep that in mind before spending an afternoon polishing one prompt. Our guide to picking the cheapest AI model for each task helps with that decision.

The 12 prompt engineering techniques

1. Be specific about the goal

Say exactly what you want, who it is for, how long it should be and what “good” looks like. Vague prompts force the model to guess, and its guesses are average by design.

Before: “Write about email marketing.”

After: “Write a 150-word LinkedIn post for small ecommerce founders explaining one reason their welcome emails underperform. End with a question. Plain, confident tone, no hashtags.”

Anthropic lists clarity first among its techniques for a reason: it fixes more problems than anything else on this list.

2. Give context and explain why

Models perform better when they understand the purpose behind a rule. Instead of “keep it short”, say “keep it under 50 words because it will appear in a mobile push notification.” The model can then make sensible choices in situations your rule did not cover.

Useful context includes your audience, the business goal, constraints (budget, brand rules, legal limits), and what has already been tried.

3. Assign a role

Role prompting means telling the model who to be: “You are a senior B2B copy editor” or “You are a skeptical CFO reviewing this proposal.” Anthropic includes role prompting among its core techniques. A role sets vocabulary, depth and point of view in one line.

Roles work best when paired with a task and criteria. “You are a tax expert” alone adds little; “You are a tax expert. Identify the three riskiest assumptions in this plan and explain each in one sentence” adds a lot.

4. Use delimiters and structure

When a prompt mixes instructions, documents and examples, separate them clearly. OpenAI’s prompt engineering guide recommends Markdown headers and XML tags as delimiters, and Anthropic’s docs also cover XML structuring.

<instructions>
Summarize the customer feedback below into 5 themes.
For each theme give a count and one representative quote.
</instructions>

<feedback>
[paste feedback here]
</feedback>

Structure stops the model from confusing your data with your instructions, which matters even more when you paste long documents. Our explainer on the context window covers how much text models can handle at once.

5. Show examples (few-shot prompting)

Few-shot prompting means including a few examples of the input and the output you want. Both OpenAI and Anthropic recommend it. Examples communicate format, tone and edge cases faster than paragraphs of description.

Classify each review as Positive, Negative or Mixed.

Review: "Fast delivery, but the size runs small."
Label: Mixed

Review: "Exactly as pictured, will buy again."
Label: Positive

Review: "[new review]"
Label:

Use 2 to 5 varied examples. If all your examples look alike, the model may copy their surface features (length, opening words) rather than the pattern you care about.

6. Specify the output format

Tell the model exactly how to shape the answer: a table with named columns, a numbered list, JSON with specific keys, or plain text under a word limit. If the output feeds another tool or a spreadsheet, a precise format saves cleanup time.

Example: “Return a Markdown table with columns: Tool, Best for, Starting price, Free plan (Yes/No). No text before or after the table.”

7. Break tasks into steps (prompt chaining)

Complex tasks go better as a sequence of smaller prompts, where each output feeds the next. Anthropic calls this prompt chaining. For a blog post, that might be: research, then outline, then draft one section at a time, then edit. Each step is easier to check and fix.

We use this exact approach in our guide on writing a blog post with AI that ranks. Within a single prompt, you can also number the steps you want the model to follow.

8. Match the prompt to the model type

OpenAI’s guide draws a helpful distinction. Treat reasoning models like “a senior co-worker”: give them the high-level goal and let them work out the steps. Treat fast GPT-style models like “a junior coworker”: give explicit, detailed instructions.

Anthropic’s docs describe a related technique, letting the model think (chain of thought), where you ask it to reason before answering. With models that already have built-in reasoning or adjustable effort settings, you often get more by raising the effort level than by writing “think step by step”. For quick, cheap models, spelling out the steps still helps.

9. Ground answers in sources

AI models can state false facts confidently. To reduce this, give the model the source material and tell it to answer only from that material. Add an escape hatch: “If the answer is not in the documents, say so.” In writing workflows, ask the model to put [CHECK] wherever it lacks a fact rather than inventing one.

For research with live citations, a search-connected tool helps. Upload reference documents into a workspace so they stay available: our guides to Claude Projects and building a custom GPT show how.

10. Put stable instructions first

In apps and API work, keep the part of the prompt that never changes (role, rules, examples, reference docs) at the start, and put the changing part (the user’s question) at the end. OpenAI’s guide notes that developer messages are “prioritized ahead of user messages”, and recommends keeping reusable content at the start of the prompt for caching.

That second point saves real money. Prompt caching only works when the start of the prompt is identical across requests. Here is a worked example with Claude Sonnet 5.5, which costs $2 per million input tokens, $2.50 per million for a 5-minute cache write and $0.20 per million for a cache hit:

• A 10,000-token system prompt sent 1,000 times without caching: 10M tokens x $2 = $20.00 in input costs.
• With caching (assuming the cache stays warm): one write of 10,000 tokens at $2.50 per million = $0.025, plus 999 cache hits of 10,000 tokens (9.99M tokens) at $0.20 per million = about $2.00. Total: about $2.02.

That is roughly 90% off the repeated part of the prompt. See our full guide to prompt caching for the rules on each platform, and the official Claude API pricing page for current rates.

11. Ask for critique, then revise

Models are often better critics than first-draft writers. After a draft, ask the model to review it against your criteria: “List the three weakest points in this proposal from the client’s perspective, then rewrite to fix them.” This two-pass approach catches gaps, generic phrasing and logic errors.

Give the critic a specific lens (a skeptical buyer, a compliance reviewer, a beginner reader). “Make it better” produces cosmetic edits; “make it convincing to a CFO who thinks this is too expensive” produces real changes.

12. Define success criteria and test

For any prompt you will reuse, especially in a product, decide what success means and test against it. Anthropic lists three prerequisites before prompt engineering: a clear definition of success criteria, ways to test against them empirically, and a first draft prompt. OpenAI similarly recommends building evals (repeatable tests) before changing production prompts.

In practice: collect 10 to 20 real inputs, including tricky ones, write down what a good answer looks like for each, and run every prompt change against the full set. You will quickly see whether a “better” prompt actually helps or just fixes one case while breaking another.

A reusable prompt template

This template combines most of the techniques above. Copy it, fill the brackets and delete what you do not need.

<role>You are [role] helping [audience].</role>

<goal>[What you want and why it matters.]</goal>

<context>[Background, constraints, what has been tried.]</context>

<sources>[Pasted documents or data. Answer only from these.]</sources>

<examples>[1 to 3 examples of good output.]</examples>

<format>[Exact structure, length and tone.]</format>

If information is missing, write [CHECK] instead of guessing.

For ready-made prompts built on these principles, browse our ChatGPT prompts for marketing, Claude prompts for writing and AI prompts for small business owners. Image prompts follow different rules; see our Midjourney prompts with styles and parameters.

Common prompt mistakes to avoid

• Stacking contradictions: “Be detailed but brief” leaves the model guessing. Pick one, or define both with numbers.
• Negative-only instructions: “Do not be boring” is weaker than describing what you do want.
• Hiding the task at the end of a long paste: put clear instructions in their own tagged section.
• Endless chats: very long conversations drift and cost more. Start a fresh chat with a clean summary when a thread gets long.
• Trusting output blindly: always verify facts, numbers and code, however good the prompt.
• Over-engineering: a 2,000-word prompt for a simple task often performs worse than three clear sentences.

Prompt engineering in ChatGPT vs Claude vs Gemini

The core techniques transfer across all major models, but there are small differences in emphasis. OpenAI’s guide highlights developer vs user message roles and the reasoning vs GPT model distinction. Anthropic’s docs emphasize XML structuring, role prompting and prompt chaining, and point to a “Prompting best practices” guide as the living reference for its latest models. Gemini responds well to the same structured approach.

Write prompts that are clear to a smart human, and they will work on every model. Then fine-tune for the model you use most. Our comparison of ChatGPT vs Claude covers which handles long instructions and documents better, and the OpenAI prompt engineering guide is worth reading in full. If you build apps, our guide on estimating AI API costs shows how prompt length affects your bill.

Frequently asked questions

Next steps

Pick one task you do with AI every week and rewrite its prompt using techniques 1, 4, 5 and 6: a specific goal, clear structure, two examples and an exact output format. Save the result as a template and reuse it. Once that works, try chaining and self-critique on a bigger task, and if you build with the API, add caching and a small test set. Model prices and features change quickly, so confirm current rates on each vendor’s pricing page before you plan costs.