Prompt engineering is the skill of writing instructions that get an AI model to do what you want, reliably. In 2026 it looks less like a bag of tricks and more like briefing a capable new colleague: say what you want and why, give the context it cannot guess, show an example, set limits and define what done looks like.
Two things changed the craft. Reasoning models now plan on their own, so some old habits waste time and tokens. Agents now take actions, so a prompt has become a work order. Below: the principles that last, what changed, eight templates and how to fix a prompt that is not working.
- Treat a prompt as a brief for a capable colleague who knows nothing about your situation.
- Five parts do most of the work: goal, context, examples, constraints and output format.
- With reasoning models, skip “think step by step.” Describe the goal and the finish line, and use the effort setting for depth.
- Agents need a work order, not a question: scope, permissions, a definition of done and a way to check.
- In controlled tests, tips, threats and expert personas did not reliably improve accuracy.
What prompt engineering means now
Anthropic’s documentation describes the right mindset. Think of the model as “a brilliant but new employee who lacks context on your norms and workflows.” Its golden rule: show your prompt to a colleague who knows little about the task. If they would be confused, the model will be too.
Current models are trained to follow instructions precisely, so plain, specific wording now does most of the work that tricks once did.
Prompt engineering is how you word the request. Context engineering covers everything else the model sees while it works, such as documents, tool results, history and memory. Our context engineering guide covers that side.
The five parts of a good prompt
| Part | The question it answers | Example |
|---|---|---|
| Goal | What do you want, and why? | A reply that keeps a customer who is upset about a late refund |
| Context | What does the model need that it cannot guess? | The audience, the background, what you already tried |
| Examples | What does good look like? | One or two samples of the tone or format |
| Constraints | What must it do or avoid? | Length, reading level, “use only the attached data” |
| Output format | What shape should the answer take? | A table with three columns, or JSON with named fields |
Say why, not only what
Reasons beat rules. Anthropic’s example: “NEVER use ellipses” works less well than explaining that the answer will be read aloud by a text-to-speech engine that cannot pronounce them. The model generalizes from the reason, so it also handles cases your rule never mentioned.
Give the context it cannot guess
Anything you leave out, the model fills with generic defaults. That is where bland, off-target answers come from. Name the audience, the situation and the stakes, and say what you already tried.
Show an example
Examples steer format and tone more reliably than descriptions. Anthropic suggests three to five examples that are relevant, varied and clearly marked, for instance wrapped in <example> tags. For reasoning models, OpenAI suggests trying zero examples first and adding a few only if the output misses.
Say what to do, not only what to avoid
“Do not use markdown” leaves the model guessing what you want instead. “Write in flowing prose paragraphs” tells it. Your own style rubs off as well: Anthropic notes that removing markdown from a prompt can reduce markdown in the answer.
Put long material first
When a prompt includes long documents, put them at the top and your question at the end. Anthropic reports that a question placed at the end can improve response quality “by up to 30 percent in tests,” especially with several documents. Separate instructions from material with headings or XML tags, as both Anthropic and OpenAI advise. The same document-first structure powers a deep research brief, covered in our guide to AI research tools.
Here is the difference in practice. Write a product description for our water bottle. leaves every real decision to the model. This version makes them:
Goal: Write a product description for our insulated steel water bottle. It goes on our online shop's product page, and it should turn browsers into buyers. Audience: Commuters and gym-goers who have had a bottle leak in their bag. Facts: 750 ml. Keeps drinks cold for 24 hours. Leakproof lid with a lock. Dishwasher safe. $34. Tone: Plain and confident, a little warm. No hype words such as "ultimate" or "best ever." Format: A headline under 8 words, two short paragraphs, then 4 bullet points of specs. Constraints: Use only the facts above. If a buyer would need something that is missing, list it as a question at the end.
What changed with reasoning models
Reasoning models work through a problem internally before they answer. On some current models, such as Claude Opus 5.5, that thinking is always on. Several old habits now waste effort or backfire.
| Old habit | Why it stopped helping | Do this now |
|---|---|---|
| “Let’s think step by step” | The model already reasons, so the phrase adds time and tokens | State the goal and what a good answer must include |
| A long hand-written procedure | The model’s own plan is often better | Give general direction, such as “think it through thoroughly” |
| “CRITICAL” and “MUST” in capitals | Newer models take instructions seriously and over-apply shouted rules | Write calm, normal instructions |
| “Double-check everything” | Some models now verify on their own, so extra checks waste time | Ask for a check against named criteria, where it matters |
| Many examples up front | The model may copy them too closely | Start with none, then add a few varied ones if needed |
| Rewording prompts for more depth | Depth is now a setting | Use the model’s effort or reasoning setting |
The evidence is consistent. OpenAI’s guidance for its reasoning models says to “avoid chain-of-thought prompts.” A 2025 Wharton study found such prompts gave reasoning models “only marginal, if any, gains” while significantly increasing time and tokens. Anthropic goes further: a prompt like “think thoroughly” often produces better reasoning than a hand-written step-by-step plan.
Prompts have also become model-specific. OpenAI notes that GPT-6 Astra runs tests and checks its work without being told, so instructions written for earlier models cause unnecessary testing. Anthropic tells developers to re-check a technique against their own tests before moving it to another model.
Prompting agents: from questions to work orders
An agent does not just answer. It runs commands, edits files, browses and calls tools, often over many steps. So its prompt needs what you would give a contractor:
The outcome, and how to know it is done. “The signup form rejects invalid emails, and all tests pass” beats “fix the form.”
Scope. Which files or systems it may touch, and which it may not.
Permissions. Anthropic’s suggested wording lets the agent take “local, reversible actions” freely, but ask before anything destructive, hard to undo or visible to other people.
Action, not suggestion. Models take wording literally. Anthropic notes that “Can you suggest some changes?” may get suggestions, while “Change this function to improve its performance” gets changes.
A way to check. A test suite, a command to run or a page to load.
One source of truth. Standing rules often live in project files such as AGENTS.md. OpenAI’s guidance for GPT-6 says your instructions outrank a skill’s, and recommends auditing instruction files for conflicts.
Agents also read content you did not write: web pages, emails and documents. Any of it can contain instructions. Tell the agent to treat what it reads as information, not orders, and keep risky actions behind your approval. For software work, our 12 prompts for coding agents put this into practice.
Eight prompt templates you can copy
Replace the parts in square brackets. They are plain text, so they work in any chat app or API. Image generators follow the same habits with their own vocabulary, which our AI image prompts guide covers. Template 3 is the seed of a fuller approach, built out in our guide to writing emails with AI.
Goal: [what you want, and why it matters] Audience: [who will read or use the result] Context: [background you cannot expect the model to know, and what you already tried] Constraints: [length, tone, must include, must avoid] Format: [the exact shape of the answer] Done means: [how you will judge the result] If anything important is missing, ask me up to 3 questions before you start.
[paste the documents here, each under its own title] Using only the documents above, answer this question: [question] Quote the sentence that supports each point. If the documents do not contain the answer, say so instead of guessing.
Rewrite the draft below for [audience and purpose]. Match the voice of this sample of my writing: [paste a paragraph you wrote] Keep every fact and number. Cut filler. Aim for [length]. Draft: [paste the draft]
Extract these fields from the text below and return only JSON:
{"company": string, "contact_name": string, "email": string or null, "deadline": "YYYY-MM-DD" or null, "budget_usd": number or null}
Use null when a field is missing. Do not guess.
Text:
[paste the text]I need to choose between [option A], [option B] and [option C] for [purpose]. My criteria, most important first: [criterion 1], [criterion 2], [criterion 3]. Compare the options in a table against these criteria, then recommend one. State the assumptions behind your recommendation and what would change your mind.
Review the draft below as a demanding editor would, for [audience]. List the 5 most important problems, most serious first. For each one, give the passage, what is wrong and a concrete fix. Do not rewrite the whole draft, and skip the praise. Draft: [paste the draft]
Goal: [the outcome, in one or two sentences] Done means: [tests pass, the page loads, the file exists] Scope: work only in [folders or systems]. Do not touch [anything off limits]. Permissions: edit files and run tests freely. Ask me before deleting anything, installing packages, pushing code or sending anything to other people. Check: before you finish, run [test or command] and show me the result. Report: summarize what you changed, what you did not do and any assumptions you made.
Here is a prompt I use: [paste the prompt] Here is an output I did not like, and why: [paste the output and your reason] Rewrite the prompt so this does not happen again. Change as little as possible, and explain each change in one line.
How to test and improve a prompt
Anthropic’s docs list three things to have before you tune a prompt: a clear definition of success, a way to test against it and a first draft.
Write down what good looks like
Choose two to five criteria you can check, such as “under 150 words,” “cites the policy number” or “no invented facts.”
Collect test inputs
Gather 10 to 20 real examples, including awkward ones: long, vague, rude or incomplete.
Run and grade them all
Run the prompt on every example and grade each result against your criteria. One good output proves little.
Change one thing at a time
Adjust a single instruction, rerun the whole set and compare. Keep the change only if the results improve.
Re-test when the model changes
A prompt tuned for one model can misfire on the next. Rerun your set before you switch.
Why so many examples? In Wharton’s tests of threats and tips, average performance barely moved, but individual questions swung by as much as 36% one way or 35% the other. One example cannot tell you which way a change cuts.
For tasks with distinct stages, chain prompts: draft, review against criteria, then refine, each as a separate call you can inspect.
Myths and mistakes
- Stating the goal and the reason behind it
- Specific context about the audience and the situation
- A few varied examples of good output
- Clear separation between instructions and material
- A check against named criteria
- Offering tips or making threats
- “You are a world-class expert” as a quality boost
- Capital letters and “CRITICAL” everywhere
- “Think step by step” on a reasoning model
- Judging a prompt by one lucky output
Wharton’s Generative AI Labs tested the popular tricks. Threatening or tipping models brought “no meaningful overall performance improvement” on hard benchmarks. Expert personas “generally did not improve accuracy” on graduate-level questions. A short role still sets tone and focus, which is how Anthropic recommends using one.
The common mistakes are quieter:
Vague goals. “Make it better” asks the model to guess what better means.
Contradictions. “Be brief” and “cover everything” cannot both win. Say which matters more.
Kitchen-sink prompts. Rules that no longer change the output still dilute the ones that do. Delete them.
Trusting fluent output. Polished prose can still be wrong. Learn why AI makes things up and how to catch it.
FAQ
Is prompt engineering still worth learning in 2026?
Yes, though it now looks like clear writing plus testing rather than tricks. Models follow instructions more precisely than before, so a vague brief gets a precisely mediocre answer. Agents raise the stakes, because a prompt now decides what they may do.
Should I still say “think step by step”?
Not to a reasoning model, which already reasons, and OpenAI advises against it. On older, non-reasoning models it can still help a little with multi-step problems, but Wharton’s tests found the gains small and the answers less consistent.
Do role prompts like “You are an expert” help?
For tone and focus, a short role helps, and Anthropic recommends one in the system prompt. For accuracy, no: a 2025 Wharton study found expert personas did not reliably improve answers to hard factual questions. Give context and criteria instead.
- A good prompt is a brief: goal, context, examples, constraints and format.
- For reasoning models, describe the goal and set the effort level instead of scripting the steps.
- For agents, add scope, permissions, a definition of done and a check.
- Test on many examples, and change one thing at a time.
Next, see how an AGENTS.md file briefs every coding agent at once, or learn how prompt injection hijacks AI. If you are studying rather than shipping code, our guide to AI for students adapts the same habits for schoolwork.
- Prompting best practices, Anthropic docs, accessed September 2026
- Prompt engineering overview, Anthropic docs, accessed September 2026
- Reasoning best practices, OpenAI, accessed September 2026
- Model guidance for GPT-6, OpenAI, accessed September 2026
- Rethinking skills and prompts for GPT-6 Astra, OpenAI, 2026
- Prompting Science Report 2: the decreasing value of chain of thought in prompting, Meincke, Mollick, Mollick and Shapiro, June 2025
- I’ll pay you or I’ll kill you, but will you care?, Wharton Generative AI Labs, August 2025
- Prompting Science Report 4: playing pretend, expert personas don’t improve factual accuracy, Basil et al., December 2025
