A good prompt engineering cheat sheet fits on one screen. Tell the AI the task, give it the context it cannot guess, assign a role if tone matters, show one or two examples, ask for a clear format and length, and then refine the answer in follow-up messages. For hard tasks, break the work into steps. Never paste passwords, ID numbers or health records into a chatbot, and check every fact before you use it.
This page is a compact reference, not a long course. Keep it open next to ChatGPT, Claude or Gemini and copy what you need. The model notes below come from each company’s own prompting guide, checked as of October 2026.
Related reading: how to write better AI prompts, what is prompt engineering, and how to learn prompt engineering. Key terms used in this guide: few-shot prompting, role prompting, and prompt injection.
The prompt engineering cheat sheet on one screen
Each row is one technique, with a line you can type and the problem it solves.
| Technique | What to write | Example line | Use it when |
|---|---|---|---|
| Role | Who the AI should act as | “Act as a patient tutor for adults new to spreadsheets.” | Tone or expertise level matters |
| Context | Who, what, why and any limits | “This is for my neighbors’ association, most are retired.” | Answers feel generic |
| Examples (few-shot) | One to five samples of what you want | “Here are two past newsletters I liked. Match their style.” | You need a specific voice or layout |
| Step by step | Numbered steps or a request to plan first | “First list the options, then compare them, then recommend one.” | The task has several parts |
| Format constraints | Length, structure, reading level | “Answer in 5 bullet points, under 100 words, plain English.” | Replies are too long or messy |
| Iteration | Feedback on the last answer | “Good. Make point 3 more concrete and cut the last paragraph.” | The first draft is close but not right |
You do not need all six every time. A quick question needs only a clear task. A letter to your bank or a summary of a long report usually needs four or five.
What prompt engineering is and why it changes output quality
Prompt engineering is the skill of writing instructions that get useful, accurate output from an AI model. A prompt is anything you type or paste into the chat: a question, a task, a document, or a mix of these. The model can only work with what you give it, so a vague request gets a vague answer.
It helps to know the main prompt types:
- Open-ended prompts invite ideas: “What are some ways to make my garden easier to care for?” Good for brainstorming.
- Closed-ended prompts ask for a specific answer or format: “List three low-maintenance shrubs for shade, with one sentence each.” Good for decisions.
- Zero-shot prompts give no examples. Fine for simple tasks.
- Few-shot prompts include examples of the output you want. Better when style or structure matters.
Clarity in prompts is the single biggest lever. Anthropic’s prompting best practices guide suggests a simple test: show your prompt to a colleague who knows nothing about the task. If they would be confused, the model will be too.
Prompt anatomy: the six parts of a strong prompt
Think of a prompt as a short brief. Most strong prompts contain these parts, in roughly this order:
- Task. One clear verb: write, summarize, compare, explain, plan, check.
- Context. Who the output is for, why you need it, and what you already know.
- Role. Optional. “Act as a friendly travel planner” sets tone and focus.
- Input. The text, notes or data the AI should work from, clearly separated.
- Examples. Optional. A sample of the style or layout you want.
- Format and limits. Length, structure, reading level, and what to leave out.
Here is a worked example, labeled as hypothetical.
Weak prompt: “Write something about our book club meeting.”
Strong prompt: “Write a friendly email to the 12 members of my book club. The next meeting moves to the community center on 6 November because the library is closed. Remind them we are discussing chapters 5 to 9. Keep it under 120 words and end with a request to reply if they need a lift.”
The second version names the reader, the facts, the action and the length. Nothing is left for the AI to guess.
If your prompt mixes instructions with a pasted document, label each part. Anthropic recommends simple tags such as <instructions> and <document> so the model does not confuse your notes with your request. OpenAI’s prompt engineering guide also suggests Markdown headings or XML tags to separate the parts of a message.
Copy-ready prompt templates by task
Swap the words in curly braces for your own details. Each template follows the anatomy above.
Writing an email or letter
Write a {tone} email to {who it is for} about {topic}.
Key facts: {facts, dates, amounts}.
What I want them to do: {action}.
Keep it under {number} words. Use plain, polite language.
Summarizing a long document
Put the document first and your question last. Anthropic reports that placing the query at the end can improve answer quality by up to 30 percent in its tests on long, multi-document inputs.
<document>
{paste the text}
</document>
Summarize this for {reader}. Give 5 bullet points, then one
sentence on what I should do next. Quote the exact lines
that support each point so I can check them.
Learning a new topic
Act as a patient tutor. I am {your background} and new to {topic}.
Explain {concept} in plain English with one everyday example.
Then ask me two short questions to check I understood.
Do not move on until I answer.
Planning a trip, week or project
Help me plan {goal} for {dates or time frame}.
My limits: {budget range, energy, mobility, time}.
First ask me up to three questions you need answered.
Then give a day-by-day plan in a table.
Comparing options before a decision
I am choosing between {option A} and {option B} for {purpose}.
What matters most to me: {criteria in order}.
Compare them in a table against these criteria, then list
what I should verify myself before deciding.
The last line of that template matters. It asks the AI to point out where it might be wrong, which makes checking faster.
Model-specific notes: Claude, Gemini and ChatGPT
All three companies publish prompting guides for developers, but most of the advice works when you type into the regular chat app.
Claude (Anthropic). The Claude prompting guide compares Claude to a smart new employee who does not know your norms yet. It recommends explaining why an instruction matters, using 3 to 5 varied examples wrapped in <example> tags, and telling the model what to do rather than what to avoid. “Write in flowing paragraphs” works better than “do not use bullet points.” The guide also notes that a single sentence of role in the instructions can focus tone.
Gemini (Google). Google’s prompting strategies page recommends always including few-shot examples, while warning that too many can make answers copy the examples too closely. It suggests one prompt per instruction for complex jobs, chaining steps so each output feeds the next. For Gemini 3 models it advises stating the goal directly, without overly persuasive language, and using one consistent format, such as XML tags or Markdown, within a prompt.
ChatGPT (OpenAI). OpenAI’s guide describes reasoning models as a senior coworker you can give a goal to, and GPT models as a junior coworker who needs explicit instructions. In practice: for a “thinking” mode, describe the outcome and let it plan; for a fast mode, spell out the steps. The guide lists a typical order for instructions: identity, then instructions, then examples, then context.
The shared lesson is more useful than the differences. Every guide stresses clear tasks, real context, examples and a defined format.
Safety: what never goes in a prompt, and how to check the answer
Good prompting includes knowing what not to share and what not to trust.
Keep sensitive data out. Do not paste passwords, card numbers, bank details, passport or ID numbers, full health records, or children’s information into any AI chat. If you need help with a document that contains them, delete or replace those details first, for example “Customer A” and “account ending XX”.
Verify every fact. AI can state wrong dates, invented quotes and made-up links with full confidence. Ask for sources, open them, and check numbers against the original page. For health, money or legal questions, use the AI to prepare questions, then confirm with a qualified professional. In a medical emergency, call 911 or your local emergency number.
Watch for prompt injection. When an AI reads web pages, emails or files for you, hidden text inside them can try to change its instructions. The OWASP Top 10 entry on prompt injection describes this risk and notes that it is unclear whether any method fully prevents it. Practical habits: give AI agents and connectors the least access they need, keep approvals on before anything is sent, deleted, paid for or shared, and be suspicious if a summary suddenly urges you to click a link or change a setting.
Check age rules. Claude requires users to be at least 18. ChatGPT and Gemini also set minimum ages in their own terms. If you share these templates with grandchildren, check each tool’s age limit first and follow it.
Common mistakes and how to fix them
Most weak answers trace back to a handful of habits:
- Asking for “something about” a topic. Fix: name the output, such as “a 200-word introduction for a church newsletter.”
- Hiding the reader. Fix: say who will read it and what they already know.
- Stacking ten requests in one message. Fix: split the job into steps and run them one at a time.
- Only saying what to avoid. Fix: describe what you want instead.
- Accepting the first draft. Fix: reply with specific feedback, such as “shorter,” “warmer,” or “add an example for point 2.”
- Starting over after a bad answer. Fix: keep the same chat and correct one thing at a time, so the AI keeps your context.
- Trusting confident numbers. Fix: ask where each figure comes from and check the source yourself.
- Copying a template blindly. Fix: templates are starting points. Add your own facts, or the output stays generic.
A quick self-test: read your prompt aloud. If a human assistant would need to ask you three questions before starting, add those answers to the prompt.
Decision framework: which technique to reach for
When an answer disappoints, match the problem to one fix and change only that. This keeps iteration fast and shows you which technique actually helped.
| If the answer is… | Try this technique | Example follow-up |
|---|---|---|
| Too generic | Add context | “This is for retirees new to email, in the UK.” |
| Wrong tone | Add a role or an example | “Act as a warm, plain-speaking librarian.” |
| Wrong layout | Show the format | “Use this layout: title, 3 bullets, one-line tip.” |
| Too long | Set limits | “Cut to 80 words. Keep only the steps.” |
| Muddled on a big task | Break it into steps | “First outline only. We will write each part next.” |
| Possibly wrong on facts | Ground and verify | “Use only the text I pasted and quote the lines you used.” |
A short worked example, labeled as hypothetical: someone asks for “a healthy weekly meal plan” and gets a generic list. Round one adds context: two adults, one vegetarian, a modest budget, 30 minutes per meal. Round two adds format: a table with days as rows and a shopping list at the end. Round three adds a limit: no ingredient that appears only once. Three small changes turn a vague answer into something usable, without writing one giant prompt.
Next step
Pick one task you do every week, such as an email, a summary or a plan, and run it through the matching template above. Then use the decision table to improve the answer in two or three follow-ups. If you would like guided practice, Coursiv offers short, bite-sized AI lessons you take at your own pace. Explore Coursiv AI lessons and build the habit one lesson at a time.
Related guides: best ChatGPT prompts and what is chain of thought prompting.