Last updated: September 25, 2026. I re-check this guide against the official OpenAI, Anthropic, and Google docs every month.
Most prompt engineering advice online still tells you to add “let’s think step by step” to everything. That tip made sense in 2022. Today, however, OpenAI, Anthropic, and Google all say their newer reasoning models already think on their own.
So this guide starts where others stop. First, you’ll learn the prompt engineering techniques all three big AI companies agree on. Then you’ll see how to prompt the two kinds of models inside ChatGPT, Claude, and Gemini. Finally, you’ll get copyable templates, a cheat sheet, and a 7-day practice plan.
Table of Contents
What Is Prompt Engineering? (Plain-English Definition)
A prompt is the text (and sometimes images or files) you give an AI tool to tell it what you want. Prompt engineering, then, is the practice of writing and organizing those instructions so the AI gives you useful, accurate, consistently formatted results.
In other words, prompt engineering is less about secret magic words and more about writing a clear brief.
Why Your Prompt Changes the Answer
Tools like ChatGPT, Claude, and Gemini run on a large language model (LLM), an AI system trained on huge amounts of text that predicts and generates language. For the bigger picture, see my guide to what artificial intelligence is.
An LLM reads your prompt in tokens, which are small chunks of text (roughly three-quarters of a word). Its context window is the maximum amount of text it can consider at once, including your prompt, your files, and its own reply. So the model only knows what you put in front of it. As a result, a vague prompt gets a generic answer, while a detailed one gets a tailored answer.
Prompt Engineering vs Just Asking ChatGPT
Asking is typing a question and hoping. Prompt engineering, by contrast, means deciding what a great answer looks like and then describing it. For example, “write a cover letter” is asking. Meanwhile, “write a 250-word cover letter for a junior accountant role at a Chicago CPA firm, highlighting my QuickBooks experience” is prompting.
Plus, you don’t need to code. You just need to think like a good manager briefing a new hire.
Why Prompt Engineering Still Matters in 2026
AI models keep getting smarter, so you might assume prompting matters less. In practice, smarter models follow your instructions more precisely, which means a fuzzy prompt gets a precisely fuzzy answer.
Anthropic puts it well: “Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.” Its golden rule is to show your prompt to a colleague with little context, because “if they’d be confused, Claude will likely be too.”
What OpenAI, Anthropic, and Google All Agree On
Here’s the most useful thing I found while researching this guide. All three labs publish their own prompt engineering guides, and their core advice lines up almost perfectly.
| Technique | OpenAI | Anthropic | |
|---|---|---|---|
| Specific instructions | Yes | Yes | Yes |
| Give context | Yes | Yes | Yes |
| Include examples | Yes | Yes, 3-5 | Yes, always |
| Tags or headings | Markdown + XML | XML tags | Clear structure |
| State the format | Yes | Yes | Yes |
| Split big tasks | Yes | Prompt chaining | Yes |
You can read each source yourself: the OpenAI prompt engineering guide, the Anthropic prompting best practices, and Google’s Gemini prompt design strategies. Because they agree, these skills work across all three tools.
The Big 2026 Shift: Reasoning Models vs Chat Models
This is the part most prompt engineering guides still get wrong. Today, you’ll meet two kinds of AI models, and each one wants a different style of prompt.
OpenAI offers the clearest analogy: “A reasoning model is like a senior co-worker. You can give them a goal to achieve and trust them to work out the details. A GPT model is like a junior coworker. They’ll perform best with explicit instructions to create a specific output.”
What Changes With Thinking Models
For years, the go-to trick was chain-of-thought (CoT) prompting, which means asking the model to work through a problem step by step before answering. A 2022 research paper by Wei and colleagues showed it helped, and the phrase “Let’s think step by step” came from another 2022 study.
Now, though, the official advice has flipped for reasoning models:
- OpenAI says prompting them to “think step by step” is “unnecessary” and advises you to “try zero shot first, then few shot if needed.”
- Anthropic says a prompt like “think thoroughly” often beats “a hand-written step-by-step plan.”
- Google says Gemini 2.5 and 3 series models “automatically generate internal ‘thinking’ text,” so asking for a plan is “generally not necessary.”
As of September 2026, many flagship models think by default. For example, Anthropic says thinking is always on for Claude Opus 5.5. OpenAI’s reasoning best practices page has the full details.
- Reasoning model? Give a clear goal, context, constraints, and what “done” looks like. Then get out of the way.
- Fast chat model? Spell out the steps, the format, and an example. Chain-of-thought still helps here.
The Same Task, Prompted Both Ways
- Step 1: list Denver neighborhoods. Step 2: pick hotels. Step 3: find hikes… (ten steps)
- Tells the model how to think, with no finish line.
- Plan a 5-day Denver trip for two, under $1,500 total, excluding flights. Include one day hike.
- Return a day-by-day table with costs. Done when the total is under budget.
Not sure which model your app uses? My ChatGPT vs Claude vs Gemini comparison explains each lineup and where the “thinking” options live. And if you’re still picking a tool for a specific job, start with my best AI tools by use case roundup.
The Building Blocks of a Good Prompt
If you’ve read my guide on how to use ChatGPT, you already know the core formula: Role + Context + Task + Format. Tell the AI who to be, explain your situation, give one clear action, and say how to deliver the answer. I won’t re-teach it here, so start there if it’s new to you.
Instead, let’s add three building blocks that turn a decent prompt into a reliable one.
For example, “keep it under 100 words” is fine. However, “keep it under 100 words because people read it on their phones” is better, because now the AI also knows to lead with the key point.
8 Core Prompt Engineering Techniques (With Examples)
These eight prompt engineering techniques cover almost everything a beginner needs. Each one comes from the official vendor guides and works in ChatGPT, Claude, and Gemini.
1. Be Specific and Say What to Do
Vague prompts get vague answers, so name the audience, length, tone, and goal. Also, phrase instructions positively. Anthropic’s example: instead of “Do not use markdown,” say “Your response should be composed of smoothly flowing prose paragraphs.”
2. Few-Shot Prompting: Show, Don’t Just Tell
Zero-shot prompting means asking for a task with instructions only. Few-shot prompting, on the other hand, means including a few examples of input and the output you want, so the model copies the pattern.
Google says few-shot prompts help control “the formatting, phrasing, scoping, or general patterning” of responses. Anthropic recommends three to five varied examples wrapped in example tags. Otherwise, the model may copy one example too closely.
Write a product description in our brand voice. Match these examples. <example> Product: Cedar candle Description: Smells like a cabin weekend. Burns 40 hours. Made in Vermont. </example> <example> Product: Wool socks Description: Warm enough for Minnesota. Soft enough for bed. </example> Now write one for: Product: [YOUR PRODUCT]
3. Chain-of-Thought (Only on Chat Models)
On a fast chat model, step-by-step reasoning still helps with math and logic. For instance, add “Work through this step by step, then give your final answer on the last line.” On a reasoning model, however, skip it, since it already reasons internally.
4. Use Delimiters and XML Tags
Delimiters are markers like triple quotes, ### headings, or XML tags such as <document> that separate the parts of a prompt. As a result, the model won’t confuse your instructions with the text you pasted.
Anthropic says “XML tags help Claude parse complex prompts unambiguously,” and you can invent any tag names you like. Plus, with long documents, put your question at the end. In Anthropic’s tests, doing so improved response quality “by up to 30 percent,” especially with complex, multi-document inputs.
5. Ask for Structured Output
Structured output means asking for a fixed format, such as a table, a list, or JSON (a simple data format developers use). For example: “Return a table with columns Tool, Monthly price, and Best for. Sort by price.” Naming the columns gives you answers you can paste straight into a spreadsheet.
6. Break Big Jobs Into a Prompt Chain
Prompt chaining means splitting a big task into a sequence of smaller prompts, where each output feeds the next. A typical chain goes outline, then draft, then critique, then final edit. For a real example, see a prompt-chaining workflow in Claude that turns a Twitter thread into a carousel.
7. Make the AI Check Its Own Work
Self-critique means asking the model to review its own answer against criteria and fix problems. Anthropic suggests adding “Before you finish, verify your answer against [test criteria],” and says it “catches errors reliably, especially for coding and math.”
8. Iterate: Treat the First Answer as a Draft
Nobody writes the perfect prompt on the first try. So read the answer, spot what’s off, and reply with a specific fix, such as “Now cut it by a third and drop the jargon.” After two or three rounds, save the final version as a template.
A hallucination is when an AI confidently states something false or made up. No prompt fully prevents it, so for money, health, or legal facts, ask for sources and verify them on an official site like IRS.gov.
Before-and-After Prompt Examples
Often, seeing the difference beats reading about it. Here are two quick prompt engineering examples, each with a note on why the “after” version wins.
- Write an email to my landlord about the broken heater.
- No facts, no tone, no goal.
- Write a polite but firm email to my landlord, Mr. Diaz. Our heater broke Monday and it’s 55 degrees inside.
- I want a repair date by Friday. Under 120 words, with a subject line.
Why it’s better: the AI now knows the facts, the tone, the deadline, and the outcome you want.
- Summarize this: [pasted text]
- The model may blend your request with the article.
<article>[pasted text]</article>Summarize the article above in 3 bullets for a busy manager.- Tags separate content from instructions, and the question comes last.
Why it’s better: the tags remove confusion, and the question sits at the end, where Anthropic’s tests show it works best.
System Prompts and Custom Instructions
A system prompt is a set of background instructions, given before the conversation starts, that shapes how the AI behaves for the whole chat. Developers write them in code, and OpenAI now calls them “developer messages” for reasoning models.
Custom instructions are the consumer-app version: a settings box that saves your preferences as a standing prompt. In ChatGPT, you’ll find them under Settings, then Personalization, and they apply to all chats. Claude and Gemini offer similar features. Anthropic notes that setting a role “focuses Claude’s behavior and tone,” and “even a single sentence makes a difference.”
About me: I'm a [JOB TITLE] in [INDUSTRY], based in [STATE]. I mostly use AI for [TOP 3 TASKS]. How to respond: Lead with the answer, then the details. Use American English and US examples. Keep answers under [NUMBER] words unless I ask for more. If my request is unclear, ask one short question first. When you're unsure of a fact, say so.
Prompting for AI Images and Video
Most prompt engineering guides skip visuals, but the same principle applies. As Google’s image docs put it, “The more specific you are, the more control you have.”
The Image Prompt Formula
Google’s official template for photorealistic images is the best starting point I’ve found.
A photorealistic [TYPE OF SHOT] of a [SUBJECT DESCRIPTION] in a [SETTING DESCRIPTION]. [DESCRIPTION OF THE LIGHT]. Shot from a [CAMERA ANGLE] with a [LENS TYPE].
For example, “a dog on a beach” becomes “A photorealistic wide shot of a golden retriever on a quiet Oregon beach at sunset. Warm, low golden light. Shot from a low angle with a 35mm lens.”
The Video Prompt Formula
Google’s Veo prompt guide says “good prompts are descriptive and clear.” It lists subject, action, style, camera position and motion, composition, focus and lens, and ambiance. As of September 2026, Veo 3.1 also makes 8-second clips with native audio, so describe sound too.
Subject: [WHO OR WHAT] Action: [WHAT HAPPENS] Style: [CINEMATIC / DOCUMENTARY / ANIMATED] Camera: [SLOW DOLLY IN / AERIAL / HANDHELD] Composition and lens: [CLOSE-UP / WIDE, SHALLOW FOCUS] Ambiance and audio: [LIGHT, MOOD, SOUNDS]
Copy-and-Paste Prompt Engineering Templates
A prompt template is a reusable prompt with blanks, like [TOPIC], that you fill in each time. These are organized by technique, not topic, so they work for any task. For finished prompts by use case, see my 50+ best ChatGPT prompts instead.
Goal: [WHAT YOU WANT TO ACHIEVE] Context: [WHO IT'S FOR, WHAT YOU ALREADY TRIED] Constraints: [BUDGET, LENGTH, DEADLINE, MUST-INCLUDES] Done when: [WHAT A FINISHED, CORRECT ANSWER LOOKS LIKE] Format: [TABLE / BULLETS / SHORT MEMO]
<document> [PASTE THE FULL TEXT HERE] </document> Based on the document above, summarize it in [NUMBER] bullets for [AUDIENCE]. Then list any deadlines, dollar amounts, or action items.
Task: [WHAT TO COMPARE OR EXTRACT] Return a table with these columns: [COLUMN 1], [COLUMN 2], [COLUMN 3]. Sort by [COLUMN]. Write "unknown" if a value is missing instead of guessing.
Step 1: Create a 5-7 section outline for [PROJECT] for [AUDIENCE]. Stop there. Step 2: Draft section [NUMBER] only, in [TONE], about [LENGTH] words. Step 3: Critique the draft as a tough editor. List the 3 biggest weaknesses, then rewrite it.
Before you finish, check your answer against: [CRITERION 1], [CRITERION 2], [CRITERION 3]. Fix anything that fails, then give me only the final version.
Technique Cheat Sheet: Which One to Use When
Bookmark this section. The first table summarizes each technique, and the second tells you which one to reach for.
Prompt Engineering Techniques at a Glance
| Technique | Try this phrase | On reasoning models? |
|---|---|---|
| Role | “You are a friendly CPA.” | Yes, one line |
| Context | “I run a bakery in Ohio.” | Yes, always |
| Few-shot | “Match these 3 examples.” | Try zero-shot first |
| Chain-of-thought | “Work step by step.” | Usually skip |
| Delimiters | “<document> … </document>“ | Yes |
| Structured output | “Return a table with…” | Yes |
| Prompt chaining | “Outline first. Stop there.” | Yes, big projects |
| Self-check | “Before you finish, verify…” | Yes |
| Constraints | “Under 150 words.” | Yes |
Which Technique When
| If you want… | Use… |
|---|---|
| The same format every time | Few-shot + structured output |
| Help with math or logic | A reasoning model |
| Answers about a long document | XML tags + question last |
| Less generic answers | More context + the why |
| A fact you can trust | Ask for sources, then verify |
| A big project done well | A prompt chain |
| Fewer errors | The self-check line |
| The same task every week | A template or custom instructions |
8 Common Prompting Mistakes (And Quick Fixes)
Honestly, even experienced users make these. Luckily, each one has a quick fix.
- One-line prompts. Fix: add the audience, the goal, and the format.
- Saying only what not to do. Fix: describe the output you do want.
- Scripting every step for a reasoning model. Fix: give a goal and a finish line.
- Using a single example. Fix: give 3-5 varied examples.
- Burying the question above a long paste. Fix: put the question last.
- Asking for everything at once. Fix: use a short prompt chain.
- Trusting facts without checking. Fix: ask for sources and verify them.
- Pasting private data. Fix: never share passwords or Social Security numbers.
Also, watch out for prompt injection, a security attack where hidden text in a document or web page tricks an AI into ignoring your instructions. So when an AI summarizes a random page, stay skeptical of odd instructions in its answer.
Is Prompt Engineering Still a Job in 2026?
Short answer: the skill is everywhere, but the standalone job title is rare.
Salary and Demand: What the Data Says
Right now, pay numbers vary widely. According to Coursera’s September 2026 salary guide, Glassdoor lists a US median total pay of $133,000 for prompt engineers. However, the same Coursera page cites ZipRecruiter’s average for “prompt engineering” at $62,977 a year. That gap shows the job category is still loosely defined.
Meanwhile, The Wall Street Journal reported in April 2025 that a Microsoft survey ranked prompt engineer second to last among new roles companies were considering. In other words, employers want the skill inside existing jobs more than a dedicated hire.
So treat prompt engineering as a career multiplier, not a career by itself. My list of 25 ways to make money with AI shows where prompting skills pay off today.
Prompt Engineering vs Context Engineering
Context engineering is the broader discipline of curating everything the model sees, including instructions, documents, tool results, and memory. Anthropic calls it “the natural progression of prompt engineering.” For everyday users, this simply means what you feed the AI matters as much as how you phrase the request.
How to Practice Prompt Engineering: A 7-Day Plan
Reading won’t make you good at this, but a week of small daily reps will.
Also, try asking the AI to improve your prompt: “Rewrite this prompt so it’s clearer and more specific, and ask me up to 3 questions first.”
Frequently Asked Questions About Prompt Engineering
What is prompt engineering in simple terms?
Prompt engineering is the skill of writing clear instructions so an AI tool gives you useful, accurate, well-formatted answers. Think of it as writing a good brief for a smart new assistant.
Is prompt engineering still relevant in 2026?
Yes, but it has changed. Newer reasoning models need less hand-holding, so step-by-step tricks matter less. However, clear goals, context, examples, and a stated format still make a big difference.
Do I need to know how to code to learn prompt engineering?
No. Everything in this guide works in the regular ChatGPT, Claude, and Gemini apps with plain English. Coding only matters if you build AI features into software.
What is the difference between zero-shot and few-shot prompting?
Zero-shot prompting gives instructions only, with no examples. Few-shot prompting adds a few examples of the output you want, so the model copies the pattern.
Do I still need chain-of-thought prompting?
Only on fast chat models. OpenAI, Anthropic, and Google all say their reasoning models already think internally, so asking them to reason step by step is usually unnecessary.
How long should a good prompt be?
As long as it needs to be. A quick question might need one sentence, while a complex task may need a few short paragraphs. If a new coworker would understand it without questions, the length is right.
Is prompt engineering a real job?
It exists, but the standalone title is rare. Coursera reports a Glassdoor median of $133,000, while ZipRecruiter shows a much lower average. Most employers want prompting as a skill inside existing roles.