Prompt library
Proven prompts,in one place.
From research, official guides and the community. Each with its source.
- 41prompts
- 8categories
- 37sources
41 prompts
Reasoning
Prompts that make the model think step by step instead of jumping to an answer.
Let’s think step by step
Makes the model write out its reasoning on maths, logic and multi-step questions instead of jumping to an answer.
Q: [your question] A: Let’s think step by step.
MultiArith accuracy · text-davinci-002 On GSM8K too: 10.4% → 40.7%.
Take a deep breath
Not written by a person: a model searching for the best prompt found it, in a Google DeepMind experiment.
Take a deep breath and work on this problem step-by-step. [Problem]
GSM8K accuracy · PaLM 2-L - Research#
Plan, then solve
The model first understands the problem and makes a plan, then carries it out: fewer skipped steps and slips.
[Problem] Let’s first understand the problem and devise a plan to solve the problem. Then, let’s carry out the plan and solve the problem step by step.
Average accuracy over 6 maths datasets, versus “Let’s think step by step”: 70.4% → 72.9% (text-davinci-003).
Take a step back
First finds the general principle behind the question, then answers the question with it. Strong on science and history.
Question: [your question] Before answering, take a step back: what is the underlying principle or concept behind this question? Explain it first, then use it to answer the question.
MMLU Chemistry accuracy · PaLM-2L - Research#
Recall similar problems
No examples needed from you: the model writes its own worked examples first, then solves the real problem.
# Problem: [problem] # Relevant problems: Recall three relevant and distinct problems. For each problem, describe it and explain the solution. # Solve the initial problem:
On MATH, versus “Let’s think step by step”: 33.9% → 37.3% (GPT-3.5-turbo).
Three experts
Squeezes the “Tree of Thoughts” method into one prompt: three experts go step by step, and whoever is wrong drops out.
Imagine three different experts are answering this question. All experts will write down 1 step of their thinking, then share it with the group. Then all experts will go on to the next step, etc. If any expert realises they’re wrong at any point then they leave. The question is... [your question]
The single-prompt version has not been widely tested. The measured result belongs to the full multi-step method: on the Game of 24, GPT-4 went from 4% to 74%.
Three paths, one answer
Solves the same problem several independent ways and keeps the answer that comes up most often.
[Problem] Solve this problem in three independent ways. Then compare the three results and give the answer that comes up most often as the final answer.
GSM8K accuracy · PaLM 540B The original method samples many separate answers and takes a majority vote; this prompt fits it into one message.
Least to most
Breaks a big problem into smaller ones and solves them in order, easiest first.
[Problem] First, break this problem down into simpler subproblems and list them. Then solve the subproblems in order, easiest first, using each answer in the next one.
SCAN accuracy · code-davinci-002 - Research#
Read the question again
As simple as giving the question twice: the model misses fewer details.
Q: [your question] Read the question again: [your question] A: Let’s think step by step.
On GSM8K, versus “Let’s think step by step”: 78.8% → 80.6% (ChatGPT).
Accuracy
Techniques that cut made-up answers and tie them to sources and checks.
Chain of verification
The model checks its own answer: a draft, check questions, independent answers, a corrected final version.
[Your question] Follow these four steps: 1. Draft an initial response. 2. Plan verification questions to fact-check the draft. 3. Answer those questions independently, without looking at the draft. 4. Generate the final verified response, fixing anything the answers contradict.
FactScore on biographies · Llama 65B Precision on list questions doubled: 0.17 → 0.36.
Permission to say “I don’t know”
The simplest guard against made-up answers: explicitly allow the model to admit uncertainty.
[Task and document] If you’re unsure about any aspect or if the [document] lacks necessary information, say “I don’t have enough information to confidently assess this.”
Anthropic: “This technique can drastically reduce false information.”
Quotes first, then answer
On long documents, grounds the answer in word-for-word quotes, so you can see where each claim comes from.
<document> [Document] </document> Find quotes from the document that are relevant to [your question]. Place these in <quotes> tags. Then, based on these quotes, answer the question. Place your answer in <answer> tags.
- Research#
Answer “according to…”
Steers the answer towards a trusted source; the model quotes more of what it really read.
[Your question] Respond to this question using only information that can be attributed to [Wikipedia].
Word-for-word quoting score (QUIP) on TriviaQA: 31.6 → 35.7 (ChatGPT).
- Research#
Clean the context first
Strips the steering and opinion out of a question, so the model is less inclined to just agree with you.
Given the following text by a user, extract the part that is unbiased and not their opinion, so that using that text alone would be good context for providing an unbiased answer to the question portion of the text. Please include the actual question or query that the user is asking. Separate this into two categories labeled with “Unbiased text context (includes all content except user’s bias):” and “Question/Query (does not include user bias/preference):”. Text by User: [original input]
Factuality on opinionated questions · LLaMA-2-70B-chat Two steps: this prompt cleans the context; you then ask the question in a fresh chat, with the clean text only.
Better questions
Prompts that sharpen the question first, or have the AI ask you.
Ask me questions first
Flips the roles: the model gathers what it needs by asking you first. One of the biggest time-savers.
From now on, I would like you to ask me questions to [achieve your goal]. Ask me the questions one at a time. When you have enough information, [create the result].
- Research#
Improve my question
Suggests a better version of each question first, so you learn what to ask.
From now on, whenever I ask a question about [topic], suggest a better version of the question to use instead and ask me if I would like to use your question instead.
- Research#
Rephrase and respond
The model restates and expands your question in its own words, then answers; a big difference on vague questions.
[Your question] Rephrase and expand the question, and respond.
Average accuracy over 10 tasks · GPT-4 - Research#
Break it into sub-questions
The model asks follow-up questions for a more accurate answer; you answer, it combines.
When I ask you a question, generate three additional questions that would help you give a more accurate answer. When I have answered the three questions, combine the answers to produce the final answers to my original question.
Turn a result into a prompt
Hand it a text, report or image you like; have it write the prompt that would produce it, and reuse that prompt.
I really like the [text / report / image] below. Write the best prompt that could have been used to produce it: describe the tone, structure, length and format, and leave the parts that change in [square brackets]. [Example]
Writing
For clearer emails, copy and translations.
Show, don’t tell
Instead of describing your style, give three examples; the model picks up tone, length and structure from them.
Below are three [emails / posts] I wrote earlier. Study my tone, sentence length and structure; then write a new one about [topic] in the same style. Example 1: [...] Example 2: [...] Example 3: [...]
TriviaQA accuracy · GPT-3 175B - Research#
Give it a role
Gives the model a point of view: the same text read through a lawyer’s, a marketer’s or a customer’s eyes.
From now on, act as [persona]. Pay close attention to [what this persona cares about]. Provide outputs that [persona] would create.
Bottom line up front
An email technique from the military: the purpose and the action needed in the first line, details after.
Rewrite the email below in BLUF (Bottom Line Up Front) form: a short opening statement that declares the purpose of the email and the action required, answering who, what, where, when and why. Then give the details as short bullet points. [Email]
English translator and improver
Write in any language; it translates and hands back richer, more polished English.
I want you to act as an English translator, spelling corrector and improver. I will speak to you in any language and you will detect the language, translate it and answer in the corrected and improved version of my text, in English. I want you to replace my simplified A0-level words and sentences with more beautiful and elegant, upper level English words and sentences. Keep the meaning same, but make them more literary. I want you to only reply the correction, the improvements and nothing else, do not write explanations. My first sentence is “istanbulu cok seviyom burada olmak cok guzel”
From the most-starred prompt collection on GitHub; even the original sample sentence is Turkish. It was started by Turkish developer Fatih Kadir Akın.
- Official guide#
Grammar fixer
Cleans up a text while keeping its meaning: spelling, punctuation, tense and word choice.
Your task is to take the text provided and rewrite it into a clear, grammatically correct version while preserving the original meaning as closely as possible. Correct any spelling mistakes, punctuation errors, verb tense issues, word choice problems, and other grammatical mistakes. [Text]
Explain like I’m five
Explains a complex topic without jargon, with an everyday comparison. Named after Reddit’s famous community.
Explain [topic] to me like I’m five: use an everyday comparison, no technical terms, five sentences at most.
Critique, then rewrite
The model critiques and improves its own draft; lifts the quality of writing, code and plans in one go.
[Task] First write a draft. Then critique the draft ruthlessly: list its weak points, gaps and mistakes. Finally, write an improved final version based on that critique.
About 20% absolute improvement on average across 7 tasks (GPT-3.5, ChatGPT and GPT-4).
Business & strategy
Ready frames for decisions, risk, meetings and spreadsheets.
Pre-mortem
Assumes the project has already failed, so you see the risks without the optimism filter.
Let’s run a pre-mortem. It is 12 months from now and [project / decision] has failed completely. Write down every plausible reason for its failure. Then pick the five most likely and most damaging, and suggest a step we can take today against each. Project: [short description]
Per the 1989 study Klein cites, imagining that an event has already happened raises the ability to identify its causes correctly by 30%.
Six thinking hats
Looks at a decision from six separate angles in turn; stops one-sided discussions.
Examine [decision] with Edward de Bono’s Six Thinking Hats. In order: White (information), Red (intuition and feelings), Black (caution), Yellow (benefits and values), Green (generating ideas), Blue (managing the thinking and concluding). Write 3–5 points per hat and end with a clear recommendation.
- Research#
What are the alternatives?
Doesn’t settle for the first idea; compares the options with their pros and cons.
Whenever I ask you to [task], if there are alternative ways to accomplish the same thing, list the best alternative approaches, compare/contrast the pros and cons of each with respect to [cost, time, risk], and include the original way that I asked. Then ask me which approach I would like to proceed with.
- Official guide#
Meeting scribe
Turns meeting notes into a short summary of topics, decisions and action items with owners.
Your task is to review the provided meeting notes and create a concise summary that captures the essential information, focusing on key takeaways and action items assigned to specific individuals or departments during the meeting. [Meeting notes]
- Official guide#
Long report to a short memo
Distils a long report into one short memo for the team: trends, risks, takeaways.
Summarize this annual report in a concise and clear manner, and identify key market trends and takeaways. Output your findings as a short memo I can send to my team. [Report]
- Official guide#
Excel formula expert
Describe the calculation you need; get the advanced Excel formula and how it works.
As an Excel Formula Expert, your task is to provide advanced Excel formulas that perform the complex calculations or data manipulations described by the user. If the user does not provide this information, ask the user to describe the desired outcome. [The calculation you need]
Say which Excel version and language you use, so the function names match.
Learning
Turn the AI into a tutor, a sparring partner or an interviewer.
AI tutor
Teaches by asking questions instead of handing over answers; used with students in Wharton classes.
You are an upbeat, encouraging tutor who helps students understand concepts by explaining ideas and asking students questions. Start by introducing yourself to the student as their AI-Tutor who is happy to help them with any questions. Only ask one question at a time. First, ask them what they would like to learn about. Wait for the response.
The full prompt is longer; this is its opening. The authors keep an updated version at moreusefulthings.com.
- Official guide#
Socratic sparring partner
Asks probing questions instead of agreeing with you, so you see your own assumptions.
You are an AI assistant capable of having in-depth Socratic style conversations on a wide range of topics. Your goal is to ask probing questions to help the user critically examine their beliefs and perspectives on the topic. Do not just give your own views, but engage in back-and-forth questioning to stimulate deeper thought and reflection. Topic: [topic]
Job interviewer
Asks one question at a time and waits for your answer, like a real interview; the best rehearsal.
I want you to act as an interviewer. I will be the candidate and you will ask me the interview questions for the [position] position. I want you to only reply as the interviewer. Do not write all the conversation at once. I want you to only do the interview with me. Ask me the questions and wait for my answers. Do not write explanations. Ask me the questions one by one like an interviewer does and wait for my answers. My first sentence is “Hi”
Code & agents
For assistants that write code and work on their own.
Three reminders for agents
For the system prompt of an assistant that works on its own: keep going until done, don’t guess, plan.
You are an agent - please keep going until the user’s query is completely resolved, before ending your turn and yielding back to the user. Only terminate your turn when you are sure that the problem is solved. If you are not sure about file content or codebase structure pertaining to the user’s request, use your tools to read files and gather the relevant information: do NOT guess or make up an answer. You MUST plan extensively before each function call, and reflect extensively on the outcomes of the previous function calls. DO NOT do this entire process by making function calls only, as this can impair your ability to solve the problem and think insightfully.
OpenAI: these three instructions raised its internal SWE-bench Verified score by close to 20%; explicit planning alone added 4%.
Explore, plan, code, commit
Stops a coding assistant from jumping straight to code: read first, then plan.
1. Read [the relevant files]; don’t write any code just yet. 2. Make a plan for how to approach this problem and show it to me. 3. Once the plan is approved, implement the solution in code. 4. Commit the result and create a pull request.
Anthropic: “Steps #1-#2 are crucial—without them, Claude tends to jump straight to coding a solution.” The live post has since changed; the original is archived.
- Official guide#
Go beyond the basics
Newer models do exactly what is asked; if you want more, say so explicitly.
Create [an analytics dashboard]. Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation.
Act as a Linux terminal
The internet’s best-known “act as” prompt; for trying and learning the command line safely.
I want you to act as a linux terminal. I will type commands and you will reply with what the terminal should show. I want you to only reply with the terminal output inside one unique code block, and nothing else. do not write explanations. do not type commands unless I instruct you to do so. when i need to tell you something in english, i will do so by putting text inside curly brackets {like this}. my first command is [pwd]
Magic phrases
Widely shared “booster” phrases, and what the evidence says today.
“This is very important to my career”
The 2023 viral phrase that adds emotional weight to a prompt.
[Task] This is very important to my career.
Relative improvement on 2023 models: 8% on Instruction Induction, 115% on BIG-Bench.
Counter-evidence: In 2025 this phrase was tested again on 5 current models (GPQA, MMLU-Pro): “boosters” like threats and tips generally had no significant effect. Meincke et al., Wharton · 2025
“I’m going to tip you”
Promising a tip for a better solution: the most talked-about item on a list of 26 “principles”.
[Task] I’m going to tip [$300K] for a better solution!
With all 26 principles applied together, GPT-4’s response quality rose 57.7% and accuracy 36.4% on average. Not the tip’s effect alone.
Counter-evidence: $1,000 and “a trillion dollars” tips were tested in 2025 on 5 models; generally no significant effect. Meincke et al., Wharton · 2025
No prompt matches this search.
How we chose
Popular is not enough.
Evidence
A peer-reviewed study, or the AI company’s own guide.
Reach
Prompts that spread online and that thousands of people use.
Source
Every card names the original source, the year and any measured effect.
Models change fast: what a phrase did yesterday may differ today. Figures belong to the model and dataset in the source; for work that matters, try two versions and compare. Sources checked on November 6, 2026.
Let me teach these to your team.
In corporate training we practise these prompts in person, on your team’s own work.
abdullah@fastemy.comThe price is set in an offer tailored to what you need.