Prompt engineering guide: how to write prompts that get useful answers

Why most prompts fail

Most prompts fail for a boring reason: they leave out the things a human colleague would need to do the job. "Write a summary of this report" tells the AI nothing about who reads the summary, how long it should be, or what decision it feeds. So the model guesses, and it guesses generic.

Prompt engineering is the habit of not making the model guess. It is closer to writing a good brief for a contractor than to programming. If you can write a clear email delegating a task, you can write a good prompt.

This guide covers the six parts of a strong prompt, how to iterate, the two techniques worth learning by name (few-shot and chain-of-thought), the one habit that fixes thin inputs ("ask me clarifying questions first"), the mistakes that cause most bad output, and a template you can copy.

The anatomy of a good prompt

You will not need every part every time. Quick questions need a sentence. Anything you would hand to a junior colleague deserves all six.

Role

Tell the model who it is for this task. "You are an experienced HR business partner at a 200-person software company" pulls in a different vocabulary and different defaults than "You are a copywriter for a consumer brand."

Keep roles specific and plausible. "You are the world's best marketer" adds nothing. "You are a B2B demand-generation manager who writes for skeptical IT buyers" changes the output.

Task

State the verb and the deliverable in one sentence. Draft, rewrite, summarize, compare, extract, critique, list, rank. "Draft a two-paragraph reply that declines the request but keeps the relationship warm" is a task. "Help me with this email" is not.

If the job has steps, number them. Models follow numbered steps far more reliably than a paragraph that implies an order.

Context

This is the part people skip, and it is the part that matters most. Context is everything the model cannot know: who the audience is, what happened before, what you have already tried, what the output is for, and any source material.

Paste the actual thing. The email thread, the meeting notes, the product description, the spreadsheet headers. A prompt with real material attached will beat a beautifully worded prompt with none.

Format

Say what the output should look like: length, structure, headings, bullets versus prose, a table with named columns, a specific number of options. "Reply in under 120 words, as three short paragraphs, no bullet points" removes a whole category of rework.

If the output will be pasted somewhere, say where. "This goes into a Slack message" and "this goes into a board memo" produce very different drafts.

Constraints

Constraints are the rules: tone, reading level, words to avoid, things not to mention, what to do when information is missing. "Do not invent figures. If a number is not in the source, write [NEEDS DATA]" is one of the most valuable sentences you can add to any prompt.

Negative constraints work best when they are specific. "Don't be salesy" is vague. "No exclamation marks, no rhetorical questions, no phrases like 'I hope this finds you well'" is enforceable.

Examples

An example of the output you want is worth more than any adjective. If you have a past email you were proud of, a summary format your boss liked, or a product description in the right voice, paste it and say "match this style and structure." That is few-shot prompting, and there is a full section on it below.

Before and after: one prompt, rewritten

Here is a prompt most people would type:

Write a LinkedIn post announcing our new scheduling feature.

What comes back is predictable: an opener about being "thrilled to announce," three generic benefits, a string of hashtags. Not wrong, just unusable.

Here is the same request with all six parts:

You are a marketing manager at a small SaaS company that sells scheduling software to independent physical therapy clinics. Our customers are clinic owners who are busy, skeptical of hype, and care about no-shows and front-desk time.

Task: draft a LinkedIn post announcing our new feature, automated appointment reminders by text.

Context: the feature sends a reminder 48 hours and 2 hours before each appointment, lets the patient confirm or reschedule with one tap, and took us six months to build because we wanted it to work with every phone. Early users told us their front desk stopped spending the first hour of the day on reminder calls.

Format: 90 to 130 words. First line must work as a standalone hook. No hashtags. End with one plain question to invite comments.

Constraints: no exclamation marks. Do not use the words "excited," "thrilled," "delighted," or "innovative." Do not invent statistics; the only claim you may make about results is the front-desk anecdote above.

Here is a past post in the voice I want:
[PASTE ONE PAST POST]

The second prompt is longer, but it takes two minutes to write and the output is usable on the first or second try. The first prompt is faster to type and produces something you will rewrite from scratch.

Notice what changed. The model now knows the audience (clinic owners), the real detail (48 hours and 2 hours, one-tap reschedule), what not to say (invented stats), and what good sounds like (your past post).

Iteration: the second prompt matters more than the first

The first answer is a draft, not a verdict. The professionals who get the most from AI treat the chat as a working session, not a vending machine.

Useful follow-ups are specific about what to change and what to keep:

• "Keep paragraphs 1 and 3 exactly as they are. Rewrite paragraph 2 to lead with the cost, not the timeline."
• "Too formal. Make it sound like a message to a colleague you like, not a client."
• "Give me three versions of the opening line only."
• "You dropped the deadline from the source. Add it back and flag anything else you left out."

Two habits help. First, when a thread gets long and the model starts drifting, restate the brief in one message rather than piling on corrections. Second, when you finally get an output you like, ask "what in my original prompt would you change to get this result on the first try?" and save the answer. That is how a personal prompt library gets built.

Few-shot prompting: show it what good looks like

Few-shot prompting means including examples of the input and output pairing you want before you give the real input. Zero-shot means you only describe the task. For anything with a house style, a fixed structure, or a judgment call, few-shot wins.

A practical version for work:

I am going to give you customer support tickets. Classify each one as BILLING, BUG, FEATURE REQUEST, or OTHER, and write a one-line summary.

Example 1
Ticket: "I was charged twice this month and the second charge has no invoice."
Output: BILLING | Duplicate charge, missing invoice for second charge.

Example 2
Ticket: "It would be great if the export button also did PDF."
Output: FEATURE REQUEST | Wants PDF export alongside existing export.

Now classify these:
[PASTE TICKETS]

Two or three examples are usually enough. Choose examples that cover the edge cases you care about, and make sure they are correct, because the model will copy your mistakes with the same confidence it copies your good choices.

Few-shot also works for voice. Paste three emails you actually sent and say "write the next one in the same voice." The guide on AI for email and writing covers the full method.

Chain-of-thought and reasoning models

Chain-of-thought prompting asks the model to work through a problem step by step before giving the answer. On multi-step problems, planning questions, and anything with arithmetic, it reduces the "confidently wrong" answers you get when the model jumps straight to a conclusion.

The simplest version is one line: "Think through this step by step, show your reasoning, then give your final answer." A stronger version tells it which steps: "First list the assumptions. Then check each one against the data I gave you. Then recommend."

One update for 2026: the major assistants now offer reasoning modes (extended thinking, "think longer," and similar labels) that do this internally. With those models, a long "think step by step" instruction adds little, and the better use of your words is to define what a good answer looks like and what it should check. Where you still want visible reasoning, say so, because the internal reasoning is often summarized or hidden.

Either way, visible reasoning is a tool for you to audit, not proof that the answer is right. Read the steps. That is where the mistakes hide.

"Ask me clarifying questions first"

The best single sentence to add to a prompt when your inputs are thin: "Before you answer, ask me up to five clarifying questions that would most change your answer. Wait for my replies."

This flips the dynamic. Instead of the model guessing what you meant and you correcting it three times, it interviews you the way a good consultant would. It works especially well for plans, proposals, job descriptions, presentations, and anything where you have context in your head that never made it into the prompt.

A variant for when you do not want to wait: "State your assumptions at the top of your answer so I can correct them." You get the draft now and a checklist of what to fix.

Common mistakes

Vague adjectives instead of specifics. "Make it engaging and professional" gives the model nothing to act on. Replace with examples, length, and words to avoid.

No format. You get a 600-word essay when you wanted five bullets. Say the shape of the output every time.

The kitchen-sink prompt. Twelve requirements in one paragraph, with no priority. Number them, and say which ones matter most if they conflict.

Asking for facts without sources. Models produce plausible-sounding numbers, citations, and quotes that do not exist. For anything factual, ask for the source, then check it. The guide on how to fact-check AI output covers the routine.

One shot, then give up. Bad first output is normal. The fix is a specific second message, not a new tab.

Forgetting the audience. Who reads this, what do they already know, and what do you want them to do? If those are not in the prompt, the output is written for nobody.

Pasting confidential data into a consumer tool. Client names, health information, salary data, unreleased financials, source code. Check your organization's policy first, and use an approved enterprise tool for anything sensitive. If in doubt, anonymize.

Treating fluency as accuracy. A confident, well-organized answer feels correct. It is not evidence of anything. Verify numbers, dates, names, and claims before they leave your hands.

A reusable prompt template

Copy this into a note and fill in the brackets. Delete any section you do not need; a short prompt with the right three parts beats a long one with filler.

ROLE
You are [ROLE, WITH THE EXPERIENCE AND POINT OF VIEW THAT FITS THIS TASK].

TASK
[ONE SENTENCE: THE VERB AND THE DELIVERABLE]. If the job has steps, do them in this order:
1. [STEP ONE]
2. [STEP TWO]
3. [STEP THREE]

CONTEXT
Audience: [WHO READS THIS AND WHAT THEY ALREADY KNOW]
Purpose: [WHAT DECISION OR ACTION THIS OUTPUT FEEDS]
Background: [WHAT HAPPENED BEFORE, WHAT I HAVE TRIED, ANYTHING YOU CANNOT KNOW]
Source material:
[PASTE THE DOCUMENT, THREAD, DATA, OR NOTES]

FORMAT
[LENGTH, STRUCTURE, HEADINGS OR BULLETS, TABLE COLUMNS, NUMBER OF OPTIONS]. This will be pasted into [WHERE IT GOES].

CONSTRAINTS
Tone: [TONE, WITH AN EXAMPLE IF POSSIBLE]
Avoid: [WORDS, PHRASES, TOPICS TO LEAVE OUT]
Do not invent facts, numbers, or quotes. If something is missing from the source, write [NEEDS DATA] rather than guessing.

EXAMPLE OF WHAT GOOD LOOKS LIKE
[PASTE A PAST OUTPUT YOU LIKED, OR WRITE "NONE"]

Before you start, ask me up to five clarifying questions that would most change your answer, then wait for my replies.

If you want this built for you interactively, the prompt builder walks through the same six parts and produces a prompt you can paste into any assistant.

Next steps

• Compare the popular acronym frameworks (RTF, CO-STAR, CRISPE, and more) in prompt frameworks that work.
• Browse ready-made prompts for writing and email and meetings and summaries.
• Read the glossary entries on few-shot prompting, chain-of-thought, and the context window.
• Learn the verification routine in how to fact-check AI output.
• Save your voice and recurring rules once with Custom GPTs, Claude Projects, or Gemini Gems.
• Ready to go deeper? The course has a full module on prompting with exercises and a quiz.