Prompts to Write Instructions for Your Custom GPT or Claude Project
A Custom GPT or Claude Project is only as good as its instructions. Most people type a line or two, then wonder why the assistant drifts, makes things up, or ignores their rules. These 10 prompts help you write instructions that hold up, fix the ones that don't, and set up a few useful assistants from scratch.
Each one started as a draft from Promptimize, then was edited by hand. Invented details came out, paste-in slots went in, and every prompt tells the AI to use only what you give it. The instructions they write cover the same five parts our Grader checks: role, context, output format, constraints, and how to reason through a request.
Want to check your finished instructions? Paste them into our free Prompt Grader for a score out of 100 and a breakdown of what's working.
They work in ChatGPT, Claude, and Gemini.
How to use these prompts
Copy a prompt into ChatGPT, Claude, or Gemini. Paste your details where marked and answer any questions it asks. Then copy the instructions it writes into your Custom GPT or Claude Project settings, and try the test messages it suggests.
Build and fix your assistant
1. Get interviewed, then get your instructions
Use it when: You want a Custom GPT or Claude Project but don't know what its instructions should say.
Why it works: It asks one question at a time, checks its summary with you, and only then writes, so the instructions come from your answers instead of guesses.
Act as an experienced designer of custom AI assistants. I want to set up my own Custom GPT or Claude Project, but I'm not sure what its instructions should say. Interview me first, then write them.
1. Ask me up to 6 short questions, one at a time, and wait for each answer. Cover what the assistant is for, who will use it, the main tasks it handles, the tone it should use, what it must never do, and what a great answer looks like. Ask a brief follow-up only if an answer is too vague to act on.
2. Summarize what you learned in 5 bullet points and ask me to confirm or correct them before you write anything.
3. Before drafting, decide which rules matter most and put them first.
4. Write the instructions as one copy-ready block in second person ("You are...", "You will..."), under these headings: Role, Context, How to work through a request, Output format, Rules.
5. After the block, list any assumptions you made, and suggest 4 example first messages I could send to try it out.
Rules: Use only what I tell you. Do not invent facts about me, my work, or my business, and do not mention files or tools I didn't bring up. Where something important is missing, ask me, or write (FILL IN: what is needed) in the instructions instead of guessing. Include a rule telling the assistant to say when it doesn't know something rather than make it up. Keep the instructions concise, in short plain sentences, with no filler like "be helpful and friendly."
What I want my assistant to help with (a sentence or two is fine):
[paste here]2. Fix an assistant that isn't behaving
Use it when: Your Custom GPT or Claude Project keeps doing something you didn't want.
Why it works: It ties each problem to the line causing it, fixes the biggest cause first, keeps what already works, and gives you test messages to confirm the fix.
Act as a troubleshooter for Custom GPT and Claude Project instructions. My assistant isn't behaving the way I want. Help me find out why and fix it. 1. Restate the gap between what it does now and what I want, in 2 sentences, so I can check you understood. 2. Trace each problem I described to the line or missing instruction most likely causing it. Quote the line when there is one. Common causes: rules that conflict, vague wording, important rules buried at the end, no example of a good answer, or instructions too long to follow reliably. 3. Rank the causes from most to least likely, and name the single change that should help most. 4. Rewrite the full instructions to fix those causes, keeping everything that already works. Use these headings: Role, Context, How to work through a request, Output format, Rules. 5. List every meaningful change you made, one line each. 6. Give me 3 test messages to send afterward to check the fix worked. Rules: Only diagnose the problems I described. Do not change the assistant's purpose, and do not remove anything without telling me. Do not add facts, policies, or details about my work that aren't in my current instructions or my notes. If the fix needs information only I have, ask me or write (FILL IN: what is needed). Keep the rewritten instructions concise. My current instructions: [paste your current instructions] What it does wrong (paste an example reply if you have one): [paste here] What I want it to do instead: [paste here]
3. Turn a prompt you keep retyping into instructions
Use it when: You paste the same long prompt into chat again and again.
Why it works: It splits your prompt into the part that never changes and the part that does, so next time you only send the new input.
Act as a prompt editor. I retype the same prompt in chat over and over. Turn it into standing instructions for a Custom GPT or Claude Project so that next time I only have to send the part that changes. 1. Sort my prompt into what stays the same every time (role, rules, tone, format) and what changes each time (the specific input). 2. Ask me what context I usually add by hand, so the instructions can include it. 3. Write the standing instructions as one copy-ready block in second person, under these headings: Role, Context, How to work through a request, Output format, Rules. Keep my original wording wherever it already works. 4. Include a rule telling the assistant to ask me for the input when my message doesn't include it, instead of guessing. 5. After the block, give me: what you kept, what you changed and why (one line each), and a one-line example of what I would type next time. Rules: Don't expand the scope or add requirements I never asked for. Don't add facts, numbers, or details that aren't in my prompt. If something is unclear, ask me before writing. Keep the instructions concise. The prompt I keep retyping: [paste your prompt]
4. Test your assistant before you rely on it
Use it when: You've set up an assistant and want to check it follows its rules before you use it or share it.
Why it works: It finds the rules most likely to fail and writes realistic test messages for each, including ones that check it admits what it doesn't know.
Act as a tester for custom AI assistants. Write a test plan to check that my Custom GPT or Claude Project actually follows its instructions before I rely on it or share it. 1. List every rule and expected behavior in my instructions, numbered. 2. Point out the rules most likely to fail: vague ones, ones that conflict, and ones a user could easily talk it out of. 3. Write 15 to 20 test messages in a table with these columns: #, Type (Normal use, Edge case, Tries to break a rule, Unclear request, Off topic), Test message, What a good reply does, Rule being tested. Write each message the way a real user would. 4. Flag any rule that can't be reliably tested, with one line on why. 5. End with up to 5 bullets on how to run the tests and what to change if one fails. Rules: Base every expected behavior on my instructions. Don't test for rules my assistant was never given, and don't include tests that need files or tools it doesn't have. Include at least 2 tests that ask for a fact it couldn't know, to check that it says so instead of making something up. My assistant's instructions: [paste your instructions] Files I added to it, if any (names and a one-line description of each): [paste here]
5. Move an assistant between ChatGPT and Claude
Use it when: You're moving a Custom GPT to a Claude Project, or the other way around.
Why it works: It keeps your rules and wording, separates behavior from settings, and gives you a checklist of what to set up by hand.
Act as an editor who adapts AI assistant instructions. I'm moving an assistant between a Custom GPT and a Claude Project, and I want it to behave the same after the move. 1. List what my instructions contain: the role, main tasks, rules, format requirements, and any mention of files, tools, or settings. 2. Flag anything that refers to a product feature or setting rather than to how the assistant should behave. Don't assume the other product works the same way. Tell me to check those in its settings myself. 3. Rewrite the instructions for the new product, keeping my wording wherever it carries over. Use these headings: Role, Context, How to work through a request, Output format, Rules. 4. Give me a checklist of things to set up by hand outside the instructions, such as files to add again. 5. List up to 3 places where the assistant might behave differently after the move. Rules: Don't drop any rule without telling me why, and don't add behaviors it didn't have before. Don't state facts about what either product can or can't do. If you're unsure about something, say so and tell me to check. Keep the instructions concise. Which direction I'm moving (Custom GPT to Claude Project, or Claude Project to Custom GPT): [paste here] My current instructions: [paste your current instructions]
Ready-made assistants to set up
6. Set up an assistant that writes in your voice
Use it when: You want drafts that sound like you, not like generic AI.
Why it works: Every style rule is backed by a quote from your own writing, and the assistant leaves a gap instead of inventing facts on your behalf.
Act as a writing editor who studies personal style. I want to set up a Custom GPT or Claude Project that drafts in my own voice. Study my writing samples, then write its instructions.
1. Read all the samples before writing anything. List the patterns you see: sentence length and rhythm, word choice, tone, humor, punctuation habits, how I open and close, and how I talk to the reader. Quote a short example from my samples for each.
2. Separate the traits that show up across several samples from one-off quirks. Keep only the consistent ones.
3. Write the instructions as one copy-ready block in second person ("You write like..."), under these headings: Role, Context, How to work through a request, Output format, Rules. Include a short "Never do" list of habits that would make the writing sound generic, and a step to check each draft against my style before replying.
4. Test it: write one short paragraph on a neutral topic in my voice, then ask me which trait still feels off.
5. Suggest which samples I should add to the assistant as reference files, and remind me to remove anything private first, such as account numbers, passwords, or other people's personal details.
Rules: Only describe traits you can point to in my samples. No vague labels like "engaging" or "authentic." In the instructions, include a rule that the assistant must never invent facts, quotes, experiences, or numbers on my behalf, and should write (FILL IN: detail) when a draft needs something it doesn't know. Keep the instructions concise. If I give you fewer than 3 samples, tell me the result will be less reliable.
What I'll use it for (emails, posts, newsletters, and so on):
[paste here]
My writing samples (3 to 5 pieces, private details removed):
[paste your writing samples]7. Set up a customer email helper for your business
Use it when: You answer the same customer questions every week and want drafts based on your real policies.
Why it works: It answers only from the policies you paste, writes a polite holding reply when it can't, and flags hard cases for you instead of promising anything.
Act as a customer support lead who sets up AI email helpers for small businesses. I want a Custom GPT or Claude Project that drafts replies to customer emails using only my real policies. I will review and send every reply myself. 1. Pull every policy rule from my text into a numbered list. Point out any that are vague or contradict each other. 2. List the kinds of customer emails my business is likely to get and which of my policies applies to each. If no policy applies, mark it as a gap. 3. Write the instructions as one copy-ready block in second person, under these headings: Role, Context, How to work through a request, Output format, Rules. Cover tone, reply structure (greeting, answer, next step, sign-off), and keeping replies short by default. 4. Include these rules in the instructions: - Answer only from the policies provided. Never promise a refund, discount, exception, delivery date, or policy that isn't written there. - When a question isn't covered, write a polite holding reply saying I'll follow up, and flag it for me at the top of the draft. - Flag angry customers, refund disputes, legal threats, and safety issues for me instead of trying to resolve them. - Be honest and kind. Don't blame the customer, make excuses, or dodge the question. 5. Write 2 sample replies using only my policies: one routine reply, and one difficult complaint handled the way the assistant should handle it, with the flag for me at the top and a polite holding reply below it. 6. End with a list of questions customers are likely to ask that my policies don't answer, so I can fill them in. Rules: Don't invent policies, prices, timelines, contact details, or business facts to fill gaps. If something is missing, write (FILL IN: what is needed). Remind me not to upload customer names, order details, or payment information as files for the assistant. My business (what we sell, in a sentence or two): [paste here] My policies (returns, shipping, refunds, hours, warranties, anything else): [paste your policies]
8. Set up a job search assistant that sticks to your resume
Use it when: You're applying to several jobs and want help tailoring each application.
Why it works: It works from a fact list built from your resume, names gaps honestly, and writes (ADD RESULT) instead of making up a number.
Act as a career coach. I want a Custom GPT or Claude Project to be my job search assistant, using only the facts in my real resume. 1. Turn my resume into a fact list: roles, dates, responsibilities, results, skills, education, and certifications. Use only what's written. If something is unclear, ask me. 2. Write the instructions as one copy-ready block in second person, under these headings: Role, Context, How to work through a request, Output format, Rules. Put the fact list in the Context section. The assistant's tasks: compare me to a job posting, tailor resume bullets, draft cover letters, and help me prepare for interviews. 3. Include a workflow for when I paste a job description: list its main requirements, match each one to facts from my resume, name the gaps, then ask me what I want written. 4. Include these rules in the instructions: - Use only facts from my resume and what I tell it. It may reword, reorder, and emphasize, but never invent skills, titles, employers, dates, numbers, or credentials. - When a result has no number in my resume, write (ADD RESULT) instead of making one up. - When a job asks for something I don't have, say so plainly and suggest honest ways to address it. Never cover a gap with invented details. - Keep the tone confident and specific, not inflated. 5. After the block, suggest 4 example first messages I could send it. Rules: Don't add skills or experience that aren't in my resume. Tell me to also add my resume to the assistant as a file. Remind me not to upload ID numbers, passwords, or other sensitive documents as files for the assistant. My resume (remove your phone number, home address, and any ID numbers first): [paste your resume]
9. Set up a study tutor that quizzes you
Use it when: You want to actually learn the material, not just get the answer.
Why it works: It asks one question at a time, gives hints before answers, and ends each session with what to review next.
Act as an experienced tutor. I want a Custom GPT or Claude Project that helps me study by quizzing me and explaining, instead of just handing me answers. 1. Write the instructions as one copy-ready block in second person, under these headings: Role, Context, How to work through a request, Output format, Rules. 2. Build in this session flow: ask what I'm studying and how well I know it, start with a few quick check questions, then quiz me one question at a time and wait for my answer. 3. Mix question types: recalling facts, applying an idea, and explaining it in my own words. 4. Use a hint ladder when I'm wrong or stuck: a small hint, then a bigger hint, then a full explanation followed by a question to check I understood. Don't give the full answer before I've tried. 5. Make questions harder after a few right answers in a row, and come back to missed topics later in the session. 6. End each session with what I got right, what I missed, and what to review next. 7. When I add my notes or textbook pages, quiz me from them first, and say when a question goes beyond them. Include these rules in the instructions: Be encouraging but honest about wrong answers. Keep explanations short. If you're not sure a fact is correct, say so instead of guessing. Help me learn, but don't write or complete graded homework, tests, or essays for me to hand in as my own. If I ask for that, offer to walk me through it instead. After the block, suggest 4 example first messages, such as "Quiz me on..." or "Explain this, then test me." What I'm studying and my current level (optional): [paste here]
10. Set up a weekly meal planner
Use it when: You want a weekly plan and grocery list that fit your diet, budget, and schedule.
Why it works: It checks your week before planning, is told never to include foods you avoid, and leaves allergy and medical questions to your doctor or dietitian.
Act as a meal planning helper for busy households. I want a Custom GPT or Claude Project that makes me a weekly meal plan that fits my diet, budget, and schedule. 1. Turn my details into a short profile: eating style, foods to avoid (marking which are allergies), weekly grocery budget, household size, cooking skill, and time available each day. If something is missing, ask me. 2. Write the instructions as one copy-ready block in second person, under these headings: Role, Context, How to work through a request, Output format, Rules. 3. Weekly workflow: first ask about changes to my schedule or budget, plan quick meals for busy days, and use leftovers or batch cooking when time is short. Reuse ingredients across meals to cut waste. 4. Each plan includes: a 7-day table (Day, Breakfast, Lunch, Dinner, Prep time), a grocery list grouped by store section, and a short prep schedule. 5. Budget: don't invent prices. If I give prices, use them to total the list. If not, suggest cheaper swaps and let me check the total against my budget. 6. Include these safety rules in the instructions: - Never include a food I've listed to avoid, even as an optional garnish. - Never say a recipe or product is safe for an allergy. Remind me to check labels for allergens and cross-contact. - For allergies, medical diets, pregnancy, or health conditions, follow what my doctor or dietitian has told me, and tell me to ask them about anything new. Don't give medical advice or make health claims. After the block, suggest 4 example first messages, such as "Plan next week. Tuesday and Thursday are busy." Rules: Use only the details I give you. Don't invent nutrition numbers or health claims. Remind me not to upload health records or medical documents as files for the assistant. My diet, foods to avoid (mark which are allergies), budget, household size, and schedule: [paste here]
FAQ
Do these work in Claude and Gemini?
Yes. Run the prompts in whichever AI you use. The instructions they write are plain text, so you can paste them into a Custom GPT, a Claude Project, or any assistant that takes custom instructions.
What should I leave out of my assistant?
Passwords, ID numbers, health records, and customer or payment details. Anything you add as instructions or files goes to your AI provider, so check your account's data settings.
My assistant still ignores a rule. What now?
Run the "Fix an assistant that isn't behaving" prompt with an example of the bad reply. Then run the test prompt to check the fix before you rely on it.
More prompt packs
Free, hand-edited ChatGPT prompts to match your resume and cover letter to any job description. Find gaps, rewrite bullets, and prep for the interview.
Free, hand-edited ChatGPT prompts for small business customer emails: angry complaints, late orders, refund requests, apologies, reviews, and more.
Free, hand-edited ChatGPT prompts to prepare for a job interview. Practice your answers, handle the salary question, run a mock interview, and follow up.
Free, hand-edited ChatGPT prompts to review a job offer, write a counteroffer, negotiate other terms, and ask for a raise. Uses only your numbers.
Free, hand-edited prompts to plan, write, test and fix your own Claude skill. Turn a task you repeat into a SKILL.md that Claude uses at the right time.
Need a prompt for something else? Describe what you're trying to do, and Promptimize writes 6 to 10 tailored versions. Free, 5 per day.
Try Promptimize