Custom instructions are the note your AI reads before it answers anything — who you are, what you do, how you want to be talked to. Most people never fill the box in, because writing it from a blank page is hard. So don't. Let it interview you and write them for you.
How: copy the prompt, paste it into ChatGPT, Claude, Copilot or Gemini, and answer its questions. It asks one or two at a time and takes about ten minutes. At the end it hands you your finished instructions, ready to paste into your settings.
# Task Your task is to **interview me and turn what you learn into high-quality Custom Instructions for the AI tools I use**, such as ChatGPT, Claude, Microsoft Copilot, Gemini, or similar assistants. The goal is to create instructions that help AI understand: * **who I am and the context I operate in** * **what I primarily use AI for** * **how I prefer AI to communicate and work with me** You should discover this through a short conversation and then write the instructions for me. **Do not ask me to write the Custom Instructions myself.** # Keep the Interview Short The entire process should take **roughly 10 minutes**. Aim for around **6–10 high-value questions in total**, asked only **1–2 at a time**. Prioritise information that is likely to materially improve future AI responses. Skip minor details, adapt your questions to what you have already learned, and stop interviewing once you have enough information to create useful instructions. # Start With Existing Context Begin by asking me for any existing information that could help you understand who I am and what I do. Specifically ask whether I can provide any of the following: * a professional bio or the About section from LinkedIn * my company or organisation website * websites for projects or products I work on * my personal website, portfolio, blog or newsletter * any other relevant public links **Actively ask for links when they are likely to be useful.** For example, if I mention that I run a company, work on a product, publish writing or have a portfolio, ask for the relevant website rather than making me explain everything manually. If I provide links and you have web access, **search or open them and use what you find as context for the interview**. You may also search the web for additional public information about the organisations, projects or products I mention when doing so would help you understand my context. If a source such as LinkedIn cannot be accessed reliably, ask me to paste the relevant bio or About text instead. Do not assume that information about a company or project describes me personally. Distinguish between: * information explicitly about me * information about organisations or projects I am associated with * reasonable inferences that should be confirmed with me Use what you learn to avoid asking questions that have already been answered and make your remaining questions more specific and useful. # What to Learn Focus the interview on the following areas. ## 1. Who am I and what do I do? Learn useful, reasonably stable context such as: * my work and areas of expertise * companies, projects or organisations I am involved with * important interests * long-term goals * where I am based, when relevant * my level of expertise in subjects I frequently discuss * recurring responsibilities or constraints that genuinely affect how an AI should help me Do not collect information simply for completeness. Include it only if knowing it is likely to improve future responses. ## 2. What do I mainly use AI for? Understand the main jobs I want AI to help with, for example: * research and fact-finding * writing and editing * brainstorming * coding * learning and explanation * strategy * decision-making * planning and organisation * analysing documents or data * creative work Use this to prioritise the instructions around how I actually use AI. ## 3. How detailed should responses generally be? Determine whether I generally prefer: * **Concise:** the answer and essential reasoning * **Balanced:** enough explanation and nuance to understand the issue without going overboard * **Deep:** comprehensive exploration, alternatives and edge cases Treat this as a default rather than a rigid rule. The appropriate level of detail can still depend on the task. ## 4. How should the AI communicate with me? Explore preferences that are likely to matter regularly, such as: * preferred language * **whether I prefer British English, American English, another variety of English, or have no preference** * directness * formal vs conversational tone * warmth and enthusiasm * humour * prose vs bullet points * headings and markdown * examples and analogies * technical vocabulary * whether to lead with the conclusion or build towards it **Specifically ask me about British vs American English if I use English.** Do not ask about every possible style preference. Focus on the ones most likely to affect my experience. ## 5. How should the AI think with me? Determine how I want the assistant to engage with my ideas. For example: * challenge my assumptions when warranted * point out mistakes or weak arguments directly * provide competing explanations or perspectives * include credible unconventional or contrarian views * make useful connections across fields * suggest better approaches when my proposed approach seems suboptimal * distinguish clearly between established facts, inference and speculation * proactively surface useful ideas I may not have considered * simply answer what I ask without unnecessary tangents Work out whether I generally want the AI to behave more like a **collaborator, critic, teacher, researcher, adviser**, or some mixture. ## 6. How should research be handled? Determine useful recurring preferences such as: * when the AI should search the web * whether current information should be verified automatically when recency matters * how heavily claims should be sourced * whether I prefer primary sources where possible * whether links or citations are useful * how conflicting evidence or uncertainty should be presented ## 7. What do AI assistants do that annoys me? Ask about recurring behaviours I want to avoid. Examples might include: * unnecessary preambles * repeating my question * excessive caveats or disclaimers * generic advice * excessive agreement or praise * over-explaining simple ideas * too many headings or bullet points * asking unnecessary clarifying questions * ending every answer with follow-up questions * constantly offering additional help * excessive jargon * overly terse responses * emojis Turn my answers into **specific behavioural guidance**, rather than a giant list of prohibitions. ## 8. Is there anything unusual that should apply almost all the time? Give me an opportunity to mention any recurring preference, convention or requirement that the previous questions have missed. # Make Questions Easy to Answer Ask only **1–2 questions per turn**. Prefer high-information questions over exhaustive questionnaires. Where a preference is difficult to describe abstractly, give me **two or three contrasting examples** and ask which feels most like what I want. For example, instead of only asking: > How verbose should responses be? show me concise examples of different response styles. You can do the same for: * tone * directness * amount of structure * how strongly the AI challenges me * how proactive it should be If I say I do not care about a preference, do not include a rule for it. Notice strong preferences I express naturally during the interview, even if they were not answers to a specific question. Do not mechanically work through every category. Adapt the interview based on my answers. # Create the Instructions Once you have enough information, **stop asking questions** and create my Custom Instructions. Separate them into: ## Personal Context Stable information about me that would materially improve future responses. ## Interaction Instructions Clear instructions describing how I want an AI assistant to communicate, reason, research and work with me. Optimise for **usefulness, clarity and brevity**. Do not create the longest possible instruction set. Identify the relatively small number of persistent facts and preferences that will have the greatest effect on the quality of my interactions with AI. Prefer concrete instructions such as: > Conclusion first, then reasoning. rather than vague descriptions such as: > I value efficient communication. **Write them telegraphically.** Concrete does not mean wordy. Use short bullet fragments rather than full sentences, and cut every "I would like you to", "please ensure that" and "you should always": > British English only. not > You should only write in British English. Both instructions produce identical behaviour; one costs half the characters. Apply the same compression to Personal Context — clipped phrases, not flowing prose. Avoid: * unnecessary prompt-engineering tricks * excessive role-play * redundant instructions * rules that simply restate behaviours a capable AI already follows * temporary information * details relevant only to one project or conversation * information unlikely to materially affect responses If the target AI has separate features for things such as Memory, Personality, style settings or project-specific instructions, you may briefly point out when something would be better placed there rather than in global Custom Instructions. Keep the finished instructions under **1,500 characters** — roughly 250 words, across both sections combined. That is the tightest box you are likely to be pasted into. If I tell you my tool allows more you may go further, but treat brevity as the default rather than something to fill. Give me the character count at the end. # Final Review Show me the finished instructions and then ask: **What here feels wrong, missing, or unlike how you'd actually want the AI to behave?** Use my response to make **one final revision**. If the result runs over 1,500 characters, cut the rules a capable assistant would already follow — never the details specific to me.