Prompt Library

The ChatGPT Prompt Cheat Sheet: 27 Reusable Formulas

27 copy-paste prompts

Stop writing prompts from scratch. These 27 fill-in-the-blank templates cover 90% of what anyone asks ChatGPT — every [BRACKET] marks a slot to fill before you hit enter. They work in Claude and Gemini too.

In short: This page contains 27 copy-paste ready prompts, organized into 5 categories with a description and pro tip for each. The first 15 prompts are free instantly — no signup needed. Hand-curated and tested by the AI Academy team.

By Louis Corneloup · Founder, Techpresso

Last updated May 15, 2026·Hand-curated & tested by the AI Academy team

Core Formulas Everyone Should Know

6 prompts

Role + Task + Context + Format

1/27

You are [ROLE — e.g. a senior financial analyst at a mid-size SaaS company]. [TASK — e.g. Review the budget below and flag the three biggest risks]. Context: [WHAT THE MODEL NEEDS TO KNOW — who this is for, what decision it feeds, any constraints]. Format your answer as [FORMAT — e.g. a bulleted list, with a one-sentence summary at the top].

The master formula. Role sets expertise and vocabulary, task sets the verb, context kills generic answers, format saves you reformatting. When in doubt, start here.

💡

Pro tip: Most people fill in Role and Task and skip Context — then blame the model for generic output. Context is the slot doing the real work.

The Constraint Stack

2/27

[TASK — e.g. Write a product announcement for our new dashboard]. Constraints: maximum [LENGTH]; written for [AUDIENCE]; must include [REQUIRED ELEMENT]; must NOT include [FORBIDDEN ELEMENT — e.g. exclamation marks, the word "excited", any pricing]; reading level: [LEVEL — e.g. a busy executive skimming on a phone].

Constraints improve output more reliably than instructions do. A stack of five hard limits forces specific choices the model would otherwise hedge on.

💡

Pro tip: The "must NOT include" slot is the most underused. Banning two or three pet phrases changes the output more than a paragraph describing your desired tone.

Audience + Outcome

3/27

Explain [TOPIC] to [SPECIFIC AUDIENCE — e.g. a new sales hire with no technical background] so that they can [CONCRETE OUTCOME — e.g. answer basic customer questions about it without escalating]. Use [NUMBER] examples drawn from [THEIR WORLD — e.g. situations a salesperson actually faces]. Skip anything they do not need for that outcome.

The fix for explanations that are technically correct but useless. Defining what the reader should be able to DO afterward forces the model to select, not summarize.

💡

Pro tip: Make the outcome testable. "Understand X" produces a lecture; "be able to do X without help" produces a usable explanation.

The Few-Shot Formula

4/27

I need more items in exactly this style. Here are examples of what I want: Example 1: [PASTE EXAMPLE]. Example 2: [PASTE EXAMPLE]. Example 3: [PASTE EXAMPLE]. Now produce [NUMBER] more for [NEW SUBJECT], matching the pattern, length, and tone of my examples exactly. Do not improve on the style — match it.

Showing beats telling. Two or three examples communicate format, length, and voice better than any description of them, which is why this is the pattern pros default to.

💡

Pro tip: Your examples ARE the instruction, so curate them: one weak example in the set teaches the model the wrong pattern, and it will faithfully reproduce the flaw.

Ask Me Questions First

5/27

I want you to [TASK — e.g. write a job description for our first marketing hire]. Before you write anything, ask me up to [NUMBER, e.g. 5] questions — one at a time — about anything that would change the output significantly. Once you have what you need, produce the [DELIVERABLE]. Do not start until you have asked.

Reverses the usual flow: instead of guessing what context the model needs, you let it pull the context out of you. Best opening move for any task where you cannot articulate the brief.

💡

Pro tip: Insist on one question at a time. Asked for five at once, the model front-loads generic questions; sequentially, each question builds on your last answer.

The Context Dump

6/27

I am going to give you a lot of background, then one specific question at the end. Background: [PASTE EVERYTHING — emails, notes, docs, history, the messier the fine]. My one question: [THE SINGLE THING YOU WANT ANSWERED]. Answer only that question, using the background. If the background does not contain enough to answer, say exactly what is missing.

Modern models handle long, messy input well — the failure mode is unfocused asks, not too much context. This formula pairs maximum context with a single sharp question.

💡

Pro tip: The closing line ("say exactly what is missing") is your hallucination guard. Without it, the model fills gaps with plausible inventions instead of flagging them.

Prompts get you started. Tutorials level you up.

A growing library of 300+ hands-on AI tutorials. New tutorials added every week.

Start 7-Day Free Trial

Writing & Editing Templates

6 prompts

The Rewrite Ladder

7/27

Here is my draft: [PASTE TEXT]. Rewrite it three times at increasing intensity. Rung 1: fix only grammar and clarity — change nothing else. Rung 2: also tighten it by about 25% and improve flow. Rung 3: full rewrite for maximum impact, any structure you want, same core message. Label each rung so I can choose how much editing I accept.

Solves the all-or-nothing problem with AI editing. You get a spectrum from light touch to full rewrite and pick the rung where it still sounds like you.

💡

Pro tip: Most people find rung 2 is the keeper. If you always pick rung 3, skip the ladder and ask for the rewrite directly; if you always pick rung 1, you need a proofreader, not a model.

Tone Transfer

8/27

Rewrite the text below so the content stays identical but the tone becomes [TARGET TONE — e.g. warm but direct, like a trusted colleague delivering hard news]. Currently it sounds [CURRENT TONE — e.g. stiff and corporate]. Keep every fact, number, and commitment unchanged. Text: [PASTE TEXT].

Separates the two jobs people usually mix: what the text says and how it sounds. Locking the facts lets the model push the tone hard without drifting on substance.

💡

Pro tip: Describe tone with a person in a situation ("a coach at halftime") rather than adjectives ("motivational") — the model has far more to grab onto.

The Tightening Pass

9/27

Cut the text below by [PERCENTAGE, e.g. 30%] without losing any ideas. Remove filler phrases, redundant sentences, throat-clearing openers, and hedge words. Do not compress by summarizing — compress by deleting what was not pulling weight. Then list the five worst offenders you cut, so I learn my own patterns. Text: [PASTE TEXT].

Almost every first draft is 25-40% too long, and the model is ruthless about cuts you are too attached to make. The offender list turns each edit into a writing lesson.

💡

Pro tip: Give a percentage, not "make it concise." A number forces real cuts; "concise" gets you a token trim of maybe 5%.

The Style Match

10/27

Here are two samples of my writing: Sample 1: [PASTE 150+ WORDS]. Sample 2: [PASTE 150+ WORDS]. Study the sentence length, rhythm, vocabulary, and quirks. Now write [NEW PIECE — e.g. a LinkedIn post about our product launch] in my voice. It should pass as something I wrote on a good day — not a parody of me, and not generic AI polish.

The cure for AI text that sounds like AI. With real samples, the model imitates your actual voice — sentence rhythm and all — instead of defaulting to its house style.

💡

Pro tip: Use samples from the same genre as the new piece. Your email voice will not transfer to a blog post; give it blog writing if you want a blog post.

The Headline Multiplier

11/27

Write [NUMBER, e.g. 12] headline options for [PIECE — e.g. a blog post about remote onboarding mistakes]. Split them across three angles: a third driven by curiosity, a third by a specific number or concrete detail, a third by a bold claim. Hard limit: [CHARACTER COUNT] characters. Then mark your top two picks and give one sentence on why each would win.

Headlines are a volume game — the tenth option is routinely better than the first. Forcing three distinct angles stops all twelve from being the same idea rephrased.

💡

Pro tip: Never take option one. Pick two finalists from different angles and test them; the angle you would not have written yourself wins more often than feels comfortable.

First Draft From Bullets

12/27

Turn these rough bullets into a [FORMAT — e.g. 600-word blog section / a one-page memo / a 90-second script] for [AUDIENCE]. My bullets: [PASTE BULLETS, MESSY IS FINE]. Keep my points in a logical order (reorder if mine is wrong), add transitions, and flag any spot where you had to guess my meaning with [CHECK: your assumption] so I can verify.

The highest-leverage division of labor in writing: you supply the thinking, the model supplies the prose. The [CHECK] flags keep its guesses visible instead of buried.

💡

Pro tip: Resist the urge to clean up your bullets first. Messy bullets with real thoughts beat polished bullets with thin ones — the model handles mess; it cannot supply substance.

Analysis & Research Templates

5 prompts

Summarize + So What

13/27

Summarize the text below in [NUMBER, e.g. 5] bullet points. Then add a final section called "So what" with 2-3 sentences on what this means specifically for [YOUR ROLE/SITUATION — e.g. a product manager deciding whether to build on this platform]. Text: [PASTE TEXT OR REPORT].

A plain summary tells you what a document says; the "So what" section makes the model do the second step people actually need — implications for your specific situation.

💡

Pro tip: The more specific the role slot, the better the payoff. "For me" gets platitudes; "for a 10-person agency with two enterprise clients" gets analysis.

Steelman Both Sides

14/27

I am deciding: [DECISION OR DEBATED QUESTION]. Make the strongest honest case FOR it, as its smartest advocate would. Then make the strongest honest case AGAINST it — no strawmen on either side. Finish with: the single piece of evidence or information that would most decisively settle this, and where I might get it.

Counterprograms both your own confirmation bias and the model's agreeable streak. The closing question converts an interesting debate into a concrete next research step.

💡

Pro tip: Do not reveal which side you lean toward — the model picks up your preference from subtle wording and quietly puts a thumb on the scale.

Extract to Table

15/27

Extract every [ENTITY — e.g. customer complaint / commitment / deadline / price mentioned] from the text below into a table with columns: [COLUMN 1], [COLUMN 2], [COLUMN 3]. One row per item, no items skipped or merged. Write "unknown" in any cell the text does not answer — do not infer or fill gaps. Text: [PASTE MESSY TEXT, EMAILS, OR NOTES].

Turns unstructured slop — email threads, call notes, feedback dumps — into structured data you can sort and act on. One of the highest-ROI uses of any model.

💡

Pro tip: The "unknown, do not infer" rule is the load-bearing line. Without it, the model helpfully invents plausible values for empty cells and you will not notice until one matters.

The Assumption Audit

16/27

Here is my plan: [PASTE PLAN, STRATEGY, OR PITCH]. List every assumption it rests on — including the implicit ones I have not stated. For each: rate how load-bearing it is (does the plan survive if this is wrong?) and how confident I should actually be in it. Then name the one assumption I should pressure-test first, and a cheap way to test it this week.

Plans rarely fail at the steps — they fail at the unexamined beliefs underneath. This surfaces them in minutes and converts the riskiest one into an immediate action.

💡

Pro tip: Run this on plans you feel confident about. Confidence is precisely the signal that assumptions have stopped being visible to you.

Compare on My Criteria

17/27

Compare [OPTION A] vs [OPTION B] (add [OPTION C] if relevant) for my specific situation: [DESCRIBE — budget, team size, constraints, what you are optimizing for]. Use these criteria: [LIST 3-5 CRITERIA THAT MATTER TO YOU]. Show a table scoring each option per criterion with one line of reasoning, then give a recommendation for MY situation — not a "both are great" non-answer.

Generic comparisons optimize for the average user, who is not you. Supplying your own criteria and constraints gets a recommendation instead of a brochure.

💡

Pro tip: If you cannot list your criteria, that IS the finding — run "Ask Me Questions First" on the decision before running this template.

Like these prompts? There are full tutorials behind them.

Learn the workflows, not just the prompts. 300+ easy-to-follow tutorials inside AI Academy — and growing every week.

Try AI Academy Free

Work & Productivity Templates

5 prompts

Meeting Notes to Action Items

18/27

Below are raw meeting notes. Produce three sections: (1) Action items as a table — task, owner, deadline; if owner or deadline was not stated, write "UNASSIGNED" rather than guessing. (2) Decisions made, one line each. (3) Open questions that were raised but not resolved. Nothing else — no summary of the discussion. Notes: [PASTE NOTES OR TRANSCRIPT].

The meeting follow-up everyone intends to write and nobody does, generated in ten seconds from however rough your notes are.

💡

Pro tip: The UNASSIGNED flags are the real product — that list is exactly what to settle in your follow-up message, and it is why those tasks were going to fall through the cracks.

The Difficult Email

19/27

Write an email to [RECIPIENT + RELATIONSHIP — e.g. a long-term client who is consistently late paying]. Goal: [WHAT MUST HAPPEN — e.g. they pay the overdue invoice within 7 days]. It must also: [PRESERVE/AVOID — e.g. preserve the relationship, avoid sounding legalistic]. Tone: [TONE]. Under [NUMBER] words. Give me two versions: one more direct, one more diplomatic.

Difficult emails burn 30 minutes of agonizing over phrasing. Naming the goal AND the relationship constraint up front gets drafts that thread the needle.

💡

Pro tip: Send neither version verbatim — take the direct version's spine and the diplomatic version's opener. The blend is almost always the right email.

The Decision Memo

20/27

Help me write a one-page decision memo on: [DECISION]. Structure: the decision in one sentence; background in 3 sentences max; options considered ([LIST THEM]) with the strongest point for and against each; my recommendation ([STATE IT]) and reasoning; what would change my mind. My raw notes: [PASTE CONTEXT]. Keep it under 400 words — it should be readable in two minutes.

The format that gets decisions approved: short, structured, and honest about trade-offs. "What would change my mind" signals rigor that bullet-point asks never do.

💡

Pro tip: Write the recommendation slot yourself before prompting. If you let the model pick, you will anchor on its choice — the memo should sharpen your decision, not make it.

Diverge Then Converge

21/27

Brainstorm [NUMBER, e.g. 20] ideas for [PROBLEM OR GOAL]. Phase 1 — diverge: no filtering, include obvious ones, weird ones, and at least three that seem impractical. Number them. Phase 2 — converge: evaluate the full list against [CRITERIA — e.g. cost under $1k, shippable in 2 weeks] and pick the top 3 with one sentence of reasoning each.

Mimics how good brainstorms actually run: generate without judging, then judge without generating. Asking for both phases in one prompt keeps the wild ideas in the pool.

💡

Pro tip: The deliberately impractical three are not filler — impossible ideas routinely contain a practical core, and the converge phase is where it gets extracted.

Explain Like I Am New Here

22/27

I just joined a team and keep hearing about [SYSTEM, PROCESS, OR JARGON — e.g. "our attribution model" or "the quarterly QBR process"]. Explain what it is, why organizations have it, what typically goes wrong with it, and the 3 questions a sharp new hire should ask about OUR version of it. Assume I am smart but have zero background in [DOMAIN].

Faster than nodding along in meetings for a month. The "what typically goes wrong" section gives you pattern recognition that usually takes a year to acquire.

💡

Pro tip: Ask the three suggested questions to an actual colleague, not the model — the prompt builds your map, but only your team knows your territory.

Power-User Patterns

5 prompts

Chain-of-Thought Trigger

23/27

[PROBLEM — e.g. paste a pricing calculation, a logic puzzle, a policy question with several conditions]. Work through this step by step, showing your reasoning at each stage before moving to the next. Consider what could make each step wrong. Only after the full walkthrough, state your final answer on its own line beginning with "ANSWER:".

For math, logic, and multi-condition problems, forcing visible intermediate steps measurably reduces errors — the model catches its own slips mid-walkthrough.

💡

Pro tip: Actually read the steps before trusting the answer line. When the model goes wrong, the broken step is visible in the chain — that is the entire point of asking for it.

The Critic Loop

24/27

Do this in three passes, all in one response. Pass 1: [TASK — e.g. write a cold outreach email to a podcast host]. Pass 2: critique your own draft harshly against these criteria: [CRITERIA — e.g. would a busy person read past line one? is there one clear ask? does it sound human?]. Pass 3: rewrite it fixing every flaw you found. Show all three passes.

First drafts from any model are mediocre; its self-critiques are surprisingly sharp. Wiring draft-critique-revision into one prompt captures the improvement automatically.

💡

Pro tip: Supply the critique criteria yourself. "Critique your draft" alone yields gentle notes; named criteria with teeth yield a genuinely better pass 3.

Flip the Interview

25/27

I want to end up with [DELIVERABLE — e.g. a personal statement / a business plan outline / a difficult performance review]. Instead of drafting from what little I have told you, interview me: ask one question at a time, follow up on interesting threads, and push back where I am vague. After [NUMBER, e.g. 6-8] questions, tell me you have enough and produce the deliverable using my actual answers.

The pattern for anything where the raw material is in your head: the model is a better interviewer than you are a briefer, and deliverables built from your answers do not sound generic.

💡

Pro tip: Give real answers, not polished ones. The follow-up questions are calibrated to what you say — performing for the interviewer just gets you a deliverable about your performance.

The Output Contract

26/27

Process the input below and return ONLY [EXACT FORMAT — e.g. valid JSON with keys "name", "priority" (1-5), and "next_step"; or: a markdown table with exactly these columns]. No introduction, no explanation, no closing remarks, no code fences around it. If the input is unprocessable, return [DEFINED ERROR FORMAT — e.g. {"error": "reason"}] instead. Input: [PASTE INPUT].

The pattern for output that feeds another tool — a spreadsheet, script, or workflow. Defining both the success shape AND the failure shape makes it dependable.

💡

Pro tip: The error-case definition is what separates a contract from a request. Without it, malformed input gets you an apologetic paragraph that breaks whatever is parsing the output.

Three Versions, One Brief

27/27

Create [DELIVERABLE — e.g. a tagline / an opening paragraph / a workshop title] for [BRIEF — audience, subject, goal]. Give me three meaningfully different takes: SAFE — the solid, expected approach; BOLD — takes a real risk in angle or tone; WILD — ignores at least one convention entirely. Label each and add one line on the trade-off it makes.

One output anchors you on one direction before you have seen the option space. Three labeled risk levels show you the map first — and calibrate your own taste over time.

💡

Pro tip: Track which label you pick across uses. Consistently choosing BOLD means your briefs are too cautious — start asking for two bold and one wild instead.

Frequently Asked Questions

Role + Task + Context + Format. Tell the model who it is ("you are a senior editor"), what to do ("cut this draft by 30%"), what it needs to know (audience, goal, constraints), and how to shape the answer (table, list, word count). You rarely need all four — but when output disappoints, the missing one is usually Context.

As long as the task is ambiguous, and no longer. "Fix the grammar" needs five words. "Write something my customers will love" needs audience, product, channel, tone, and length, or you are delegating those decisions to the model's defaults. The practical rule: every detail that would change the output belongs in the prompt; everything else is noise.

Yes. All 27 templates are model-agnostic — they encode how to communicate a task clearly, which transfers across ChatGPT, Claude, Gemini, and whatever ships next. You may notice stylistic differences (Claude tends to follow long constraint lists tightly, for instance), but the structures themselves work everywhere.

Accepting the first answer. The single biggest quality gain in 2026 still comes from iteration: ask for three versions, request a self-critique and revision (the Critic Loop above), or reply with what is wrong and let the model fix it. People who treat output one as a first draft get dramatically better results than people who treat it as the answer.

The fiddly tricks aged badly; the fundamentals aged well. Models now tolerate typos and forgive awkward phrasing, so "magic words" matter less than ever. But no model can read your mind — being specific about audience, constraints, format, and success criteria is just clear delegation, and it will matter as long as you are asking anything to do work for you.

Three options, in increasing order of convenience: keep a notes file of your filled-in favorites; use ChatGPT custom instructions or a Claude project to store standing context (your role, audience, style) so every prompt starts pre-loaded; or set up text-expander snippets so typing a shortcut pastes the whole template with brackets ready to fill.

Prompts are the starting line. Tutorials are the finish.

A growing library of 300+ hands-on tutorials on ChatGPT, Claude, Midjourney, and 50+ AI tools. New tutorials added every week.

Start 7-Day Free Trial See how AI Academy teaches AI for Business

7-day free trial. Cancel anytime.

Browse All Prompt Collections