Launch price: $49 one-time for all 38 specialists. Get it →
Claude prompts
10 Claude Prompt Engineering Prompts: Generators, Improvers and Templates
Prompt generator, prompt improver, custom instructions and templates for better Claude results.
Updated October 202610 promptsFree to copy, no signup
These prompts help you write better prompts. Use them to generate a structured prompt from a rough goal, fix one that gives weak results, set up custom instructions and Projects, and keep Claude honest when accuracy matters.
How to use these prompts
Start with the prompt generator when you only have a rough goal, and the prompt improver when you already have a prompt that underperforms.
Answer the clarifying questions in short fragments; Claude needs facts, not polished writing.
Test any new prompt in a fresh chat on a real case before saving it as a template.
Save prompts that work in your custom instructions, a Claude Project or a notes file so you can reuse them in one paste.
Use it when: Use when you know what you want but not how to ask for it; Claude asks a few questions, then writes a complete, structured prompt you can reuse.
Prompt 01
I want you to act as a prompt engineer and write a strong prompt for Claude that I can reuse.
My rough goal: [DESCRIBE WHAT YOU WANT DONE IN A SENTENCE OR TWO]
Before writing anything, ask me up to 5 short questions about whatever is missing from this list:
1. The goal, and what a great result looks like
2. Who the output is for (audience and their level)
3. What inputs I will provide (documents, data, examples)
4. The output format and length
5. Success criteria, and anything that would make the result unusable
Then write the prompt with these sections:
<role> one line on the expertise needed
<context> background and why the task matters
<instructions> numbered steps in the order Claude should do them
<input> [PLACEHOLDERS] for what I will paste in each time
<output_format> exact structure, length and tone
<quality_checks> 3 to 5 checks Claude runs before answering
Phrase instructions as what to do, and add a short reason to any rule that is not obvious, so Claude can apply it sensibly. Keep the prompt as short as it can be while still complete. After the prompt, list the assumptions you made so I can correct them.
Why it works: Most weak prompts fail because the goal, audience or output format was never stated, so the five questions force those gaps into the open before any prompt is written. The fixed sections (role, context, instructions, output format, quality checks) give Claude a clear structure to fill, and the closing list of assumptions shows you exactly where the generated prompt guessed.
Tip: Paste one example of output you liked; Claude can build it into the prompt as an <example>, which usually helps more than extra rules.
2. Prompt improver: critique and rewrite my prompt
Claude or Claude Code
Use it when: Use when a prompt gives mediocre or inconsistent results; you get a diagnosis, a rewritten version, and a change log explaining each fix.
Prompt 02
Review and improve the prompt below. I use it with Claude for: [WHAT THE PROMPT IS FOR]. The problem I see in the results: [WHAT GOES WRONG, OR "NOT SURE"].
<prompt>
[PASTE YOUR PROMPT]
</prompt>
Work in this order:
1. Diagnose. List the specific weaknesses, such as missing context, an unclear goal, a vague output format, conflicting instructions, rules with no reason given, instructions phrased only as "don't", missing examples, or a task that should be split into steps. Quote the line each issue comes from.
2. Ask me up to 3 questions if a fix depends on information only I have. If nothing is blocking, skip this step.
3. Rewrite the prompt. Keep everything that already works and keep my intent. Use XML-style sections (<context>, <instructions>, <output_format>) where they make it clearer, and [PLACEHOLDERS] for inputs that change each time.
4. Show a change log: one line per change, with the reason.
Make it longer only where that adds something Claude needs. If a shorter prompt would work better, write the shorter one.
Why it works: Asking for a diagnosis with quoted lines before the rewrite makes Claude find the actual cause of bad results instead of just polishing the wording. Keeping what already works and showing a change log with reasons stops the rewrite from drifting away from your intent, and lets you reject any change you disagree with.
Tip: Include one bad output the original prompt produced, so Claude can target the actual failure instead of guessing.
3. Interview me first: questions before answers
Claude or Claude Code
Use it when: Use before any complex task (a plan, a document, a strategy) so Claude gathers the details it needs one question at a time instead of guessing.
Prompt 03
I want help with: [YOUR TASK, E.G. A LAUNCH PLAN FOR MY CATERING SERVICE]
Before you produce anything, interview me, so the result fits my real situation rather than a generic one.
How to run the interview:
- Ask one question at a time, and wait for my answer before asking the next.
- Start with the questions whose answers change the result most: goal, constraints, audience, deadline, budget, and what I have already tried.
- Where it helps, offer 2 to 4 likely options I can pick from, plus "something else".
- If an answer is vague, ask one follow-up before moving on.
- Stop when you have enough to do excellent work, usually 5 to 10 questions. Then tell me you are ready and give a 3-line summary of what you understood, so I can correct it.
Once I confirm the summary, produce [THE DELIVERABLE] in this format: [FORMAT, E.G. A ONE-PAGE PLAN WITH WEEKLY MILESTONES].
Why it works: Asking one question at a time stops Claude from front-loading a long questionnaire that you skim and half answer. Ordering questions by how much they change the result means the important facts come first, and the 3-line summary gives you a cheap checkpoint to fix misunderstandings before any real work is produced.
Tip: Answer in quick fragments; you are supplying facts, not writing, and Claude will follow up if something is unclear.
4. Pre-mortem prompt before a big task
Claude or Claude Code
Use it when: Use before handing Claude a long or high-stakes task; it predicts how it is most likely to fail and proposes an instruction to prevent each failure, then waits.
Prompt 04
Before you start on the task below, run a pre-mortem.
<task>
[DESCRIBE THE TASK AND PASTE ANY MATERIAL]
</task>
Imagine you have finished this task and I am disappointed with the result. List the 5 most likely reasons, ranked by likelihood and cost. For each, give:
1. The failure, in one concrete sentence (for example "filled gaps in the data with invented figures")
2. Why it would happen with this specific task
3. One instruction I could add to the prompt that would prevent it
4. Any information you need from me to avoid it
Then stop and wait for my reply before starting the task. I will tell you which instructions to adopt and answer your questions, and you will then do the task with those instructions in force.
Keep every risk specific to this task. A generic risk like "misunderstanding the request" only belongs on the list if you can point to the exact ambiguity.
Why it works: Imagining the task has already failed makes Claude name concrete failure modes, such as invented figures, instead of vague risks. Pairing each failure with a preventing instruction turns the list into prompt improvements you can adopt, and the explicit wait means the task only starts once those instructions are in force.
Tip: Save the prevention instructions you adopt; after a few pre-mortems on similar work you will have a solid reusable prompt.
5. Brutally honest feedback prompt for ideas
Claude or Claude Code
Use it when: Use when you need an idea, plan or decision pressure-tested rather than praised; you get a steelman, the strongest objections, and a clear verdict.
Prompt 05
I want an honest critique of my [IDEA / PLAN / DECISION], not encouragement. I am about to commit time and money to this, so automatic agreement is no use to me.
<idea>
[DESCRIBE IT: WHO IT IS FOR, WHAT IT COSTS, AND WHY YOU THINK IT WILL WORK]
</idea>
Respond in this structure:
1. Steelman: the strongest honest case for it, in 3 to 5 sentences.
2. Objections: the 5 strongest objections, ranked by how likely each is to sink it. For each, say what evidence would prove the objection wrong.
3. Hidden assumptions: what I am taking for granted without having tested it.
4. Cheapest test: the fastest, lowest-cost way to find out whether the biggest risk is real.
5. Verdict: go, change it, or drop it, with one paragraph of reasoning and your confidence level.
Leave out compliments and filler. If something is genuinely strong, say so once in the steelman and move on. Where you lack the information to judge a point, say what you would need instead of guessing.
Why it works: Stating that agreement is useless to you, and why, counters the model's tendency to be encouraging by default. The steelman-then-objections structure means the criticism is aimed at the strongest version of your idea, and asking what evidence would disprove each objection plus a cheapest test turns the critique into something you can act on.
Tip: List the objections you have already heard and answered, so Claude spends its effort on new ones.
6. Reduce hallucinations: research mode prompt
Claude or Claude Code
Use it when: Use when accuracy matters more than fluency; Claude answers only from evidence, quotes its sources, marks confidence, and says plainly when it does not know.
Prompt 06
<documents>
[PASTE SOURCE DOCUMENTS, OR WRITE "NONE: USE YOUR GENERAL KNOWLEDGE"]
</documents>
<question>
[YOUR QUESTION]
</question>
Answer the question above using careful research rules. I will rely on this, so an error costs me more than a gap.
1. If documents are provided, first pull the exact quotes relevant to the question into <quotes> tags, then answer using only those quotes, noting which quote supports each claim.
2. If the material, or knowledge you are confident in, does not support an answer, say "I don't know" or "the sources don't say". You have my permission to leave gaps.
3. Mark each key claim as high, medium or low confidence.
4. Keep facts separate from interpretation, and label inferences as inferences.
5. Give only citations, URLs, statistics and quotes that you can see in the material. If I need one that is not there, tell me where I could look for it.
6. After drafting, reread your answer and retract or flag any claim you cannot back up.
Format: a direct answer in 2 to 4 sentences, then the supporting points with their confidence, then open questions.
Why it works: Extracting exact quotes first and answering only from them grounds every claim in text you can check. Explicit permission to say "I don't know" removes the pressure to fill gaps with plausible guesses, and the final reread to retract unsupported claims catches errors that slip through the first draft.
Tip: For long files, keep the documents first and the question last, as above; Claude handles long material more accurately in that order.
7. Claude custom instructions builder
Claude chat
Use it when: Use once to write the personal preferences in Claude's settings, so every new chat already knows your role, context and preferred style.
Prompt 07
Help me write the personal preferences I will paste into Claude's settings, so every new conversation starts with the right context about me.
Interview me first, one question at a time, covering:
1. My role, field and level of expertise
2. What I mostly use Claude for
3. How I like answers: length, structure, tone, and use of headings and bullet points
4. Habits in AI answers that annoy me (for example piles of caveats, praise, or restating my question)
5. Spelling (US or UK), units, time zone, and the tools I use
6. When I want you to ask a clarifying question versus make a reasonable assumption
Then write the preferences:
- Under 200 words, in the first person ("I am...", "I prefer...")
- Specific and testable: "answer in under 150 words unless I ask for depth" rather than "be concise"
- A short reason for any preference that is not obvious, so it gets applied with judgment
- Grouped under: About me, How to respond, When to check with me
Finally, suggest 2 test questions I can ask in a new chat to confirm the preferences work.
Why it works: Interviewing you first means the preferences describe your actual work instead of generic advice like "be concise". Requiring specific, testable rules with short reasons lets Claude apply them with judgment, and the two test questions let you confirm the preferences change behavior before you rely on them.
Tip: Revisit them after a week; add one line for each thing you caught yourself correcting more than twice.
8. Claude Project instructions builder
Claude chat
Use it when: Use when you have a recurring job (weekly reports, client emails, lesson plans) and want a Claude Project whose instructions and knowledge files get it right every time.
Prompt 08
I want to set up a Claude Project for a recurring job, and I need you to write its project instructions.
The job: [DESCRIBE THE RECURRING TASK, E.G. DRAFTING WEEKLY CLIENT STATUS UPDATES]
How often I do it: [FREQUENCY]
Ask me what you need to know, one question at a time (6 questions or fewer): who the output is for, what inputs I bring each time, what a great result looks like, what usually goes wrong, tone and format, and any rules from my organization.
Then deliver three things:
1. Project instructions (150 to 400 words) with sections: Purpose, Inputs I will provide, Steps to follow, Output format, Quality checks, When to ask me instead of assuming. Give reasons for rules that are not obvious.
2. A knowledge file plan: which files I should upload (templates, style guide, past examples, glossary), what each should contain, and how the instructions refer to each file by name.
3. A starter message I paste at the start of each new chat in the project, with [PLACEHOLDERS] for the parts that change.
Keep the instructions focused on this one job; general preferences about me belong in my profile settings.
Why it works: A Project works best when its instructions describe one recurring job, so the prompt scopes everything to that job and pushes general preferences elsewhere. Planning knowledge files and referring to them by name tells Claude when to use each one, and the starter message with placeholders makes every new chat begin with the same complete inputs.
Tip: Upload two or three real examples of the output at its best as knowledge files; examples steer Claude more reliably than descriptions.
9. Turn this chat into a reusable prompt template
Claude or Claude Code
Use it when: Use at the end of a conversation that finally produced what you wanted; Claude distills it into one prompt with placeholders, so next time takes a single message.
Prompt 09
This conversation eventually produced the result I wanted, but it took several rounds of corrections. Turn it into one reusable prompt so I can get the same quality in a single message next time.
1. Review the whole conversation and list every correction, preference and clarification I gave you. These matter most, because they are exactly what the first attempt was missing.
2. Separate what was specific to this case (names, numbers, topic) from what applies every time.
3. Write the reusable prompt with these sections: <context>, <instructions>, <input>, <output_format>, <quality_checks>. Turn each case-specific detail into a [PLACEHOLDER IN CAPS] with a short hint, for example [CLIENT NAME] or [PASTE LAST MONTH'S NUMBERS].
4. If the final output we settled on makes a good example, include a shortened version in <example> tags, labeled as a guide to format rather than content to copy.
5. Under the prompt, list each placeholder with one line on what to put in it.
Keep it tight: include a rule only if it changed the result in this conversation.
Why it works: The corrections you gave during a long chat are exactly what the first attempt was missing, so listing them first captures the knowledge that made the final result good. Separating case-specific details into placeholders makes the prompt reusable, and including a rule only if it changed the result keeps the template short.
Tip: Test the template in a fresh chat on a different case; anything you still have to correct goes back into the template.
10. Think it through prompt for hard decisions
Claude or Claude Code
Use it when: Use for hard problems with no obvious answer; Claude works through observations, options and trade-offs before recommending, and is upfront about uncertainty.
Prompt 10
Help me think through a hard problem. I care more about sound reasoning than a fast or confident-sounding answer.
<problem>
[DESCRIBE THE PROBLEM OR DECISION, THE CONSTRAINTS, AND WHAT IS AT STAKE]
</problem>
Work through it under these headings:
1. Observations: what we actually know, kept separate from assumptions. Label each assumption.
2. The real question: restate what needs deciding in one sentence. If I am asking the wrong question, say so.
3. Options: 3 to 5 genuinely different options, including doing nothing and a smaller version.
4. Trade-offs: for each option, the main benefit, cost and risk, and what would have to be true for it to be the best choice.
5. Decision: your recommendation and reasoning, a confidence level (low, medium or high), and the one piece of information that would most change your mind.
6. Next step: the first concrete action, doable this week.
Where the evidence is thin, say so plainly instead of covering the gap with certainty. If something important is missing, ask me before step 5.
Why it works: Separating observations from labeled assumptions stops guesses from being treated as facts later in the reasoning. Requiring several genuinely different options, including doing nothing, prevents an early fixation on the first idea, and asking for a confidence level plus the information that would change the decision keeps the answer honest about uncertainty.
Tip: If your Claude app offers extended thinking, switch it on for this one; the headings keep the final answer easy to scan.
Getting more out of every prompt
Give context and the reason behind rules
Claude follows instructions better when it knows why they matter. A rule like "keep it under 150 words because readers see it on a phone" is applied more sensibly than a bare word limit. Treat Claude like a capable new colleague who has no background on your situation.
Show examples of the output you want
One to three examples in <example> tags often steer format and tone more reliably than paragraphs of description. Make the examples varied so Claude copies the pattern rather than the specific content. Label them as examples so they are not mistaken for input.
Use XML-style tags to separate parts
Wrapping documents, instructions and examples in tags such as <document> and <instructions> helps Claude tell them apart. This matters most when you paste long material. Put long documents near the top and your question at the end.
Say what to do, not only what to avoid
Positive instructions such as "write in short paragraphs of plain prose" tend to work better than a list of things to avoid. Calm, direct wording also works better than all-caps warnings, which can make Claude overcautious.
Frequently asked questions
How do I write good prompts for Claude?
State the goal, who the output is for, the inputs you are providing and the format you want, and explain why any important rule matters. Separate long material from instructions with XML-style tags and add an example of good output if you have one. If details are missing, ask Claude to ask you questions before it starts.
Can Claude write prompts for me?
Yes. Describe your goal and ask Claude to interview you about audience, inputs, format and success criteria, then write a structured prompt. The prompt generator on this page does exactly that and lists its assumptions so you can correct them.
Where do I put custom instructions in Claude?
In the Claude apps you can add personal preferences in your profile settings, and they apply to every conversation. For a specific recurring job, a Claude Project lets you add project instructions and knowledge files that apply only to chats inside that Project. In Claude Code, the equivalent is a CLAUDE.md file in your repository.
How do I stop Claude from making things up?
Give Claude the source material, ask it to quote the relevant passages before answering, and explicitly allow it to say it does not know. Ask it to mark confidence and to retract any claim it cannot support. These steps reduce errors a lot, but you should still verify important facts yourself.