10 Rules For Using Claude
With A Prompt For Each

10 RULES FOR USING CLAUDE: ALL TEN PROMPTS From Learn AI With Mariah (learnaiwithmariah.com/guides/10-rules-for-using-claude/) Fill in anything in [brackets], then paste the prompt into Claude. RULE 1: ASK ME QUESTIONS BEFORE YOU START Before you do any work on the task below, interview me so you don't have to guess what I want. THE TASK [Describe what you want in a sentence or two. Rough is fine.] HOW TO RUN IT 1. Start by telling me, in 2 or 3 lines, what you think I want and what you would assume if I said "just go." I'd rather correct your guesses than answer from scratch. 2. Then ask the questions whose answers would change the result the most. Ask ONE at a time. With each question, give me your best guess or 2 to 3 options to pick from, so I can answer in a few words. 3. Only ask about things you can't work out yourself. Usually that means: - Who it's for and what they'll do with it - What a great result looks like to me - Must-haves and must-nots - Length, format and where it's going - What I've already tried, and anything I'd hate 4. Stop at 5 questions unless a big gap is still open. If a detail is minor, make a sensible choice and tell me what you chose. 5. If I say "you decide," decide, and write it down as an assumption. 6. I might be talking instead of typing. Expect rambly answers and pull out what matters. WHEN YOU HAVE ENOUGH Give me a 4 to 6 line brief with the goal, the audience, the constraints and your assumptions. Ask "Good to go?" and start only when I say yes. ========== RULE 2: STUDY MY VOICE BEFORE YOU WRITE FOR ME I want you to learn how I write, so anything you draft for me sounds like me and not like AI. At the end, hand me a reusable Voice Profile I can save once and use in every chat. STEP 1: GATHER WHAT YOU CAN - First, check your memory and our past chats for anything I've written or said about how I like to write. Tell me what you found in 2 or 3 bullets, or say you found nothing. - Then ask me to paste or upload 3 to 6 real things I wrote myself, with no AI help. A good mix is a casual text or DM, a work email, a social post, something I wrote when I was excited or annoyed, and something longer. Tell me unpolished is better. If I only have 1 or 2, work with them and tell me how confident you are. STEP 2: SHOW ME YOUR READ Study the samples and describe my voice back to me as your best guess. Cover: - Sentence length and rhythm - How I open and how I close - Words and phrases I repeat, and my verbal tics - How formal or casual I am, and whether that changes by audience - Punctuation habits (exclamation points, emojis, dashes, ellipses, capital letters) - Humor, how much and what kind - How I give opinions, hedged or blunt - Things I never do Quote one of my own lines as evidence for each point. STEP 3: INTERVIEW ME, ONE QUESTION AT A TIME Ask up to 6 short questions about what samples can't show. Lead each one with your guess so I can just say "yes" or fix it. Topics to cover: - Who I write for most, and how I want them to feel after reading - 2 or 3 writers, creators or brands whose voice I love, and what I'd borrow from each - Words, phrases and habits I hate seeing in AI writing - How my voice shifts between a quick DM and something public - Things I would never say I might be dictating, so expect long, rambly answers. Pull out what matters and don't make me repeat myself. STEP 4: TEST IT ON ME Write the same short message twice on [TOPIC, or pick one from my samples]. Version A is your default voice. Version B is my voice. Ask me what's off about B. Update the profile based on my answer. Do one more round if I'm still not happy. STEP 5: HAND ME MY VOICE PROFILE Write the final profile as one clean block titled MY VOICE PROFILE, with these sections: 1. The short version. My voice in 3 sentences. 2. Do. 8 to 12 specific rules, each with a real example line from my samples. 3. Don't. 8 to 12 things I never do, including the AI phrases to ban. 4. By context. How it shifts for casual messages, work email, public posts and long pieces. 5. Before and after. One paragraph of generic AI writing, rewritten in my voice. 6. Self-check. 5 questions to run on any draft before you show it to me. Keep it under 600 words so it fits in instructions. THEN HELP ME SAVE IT - Offer to turn it into a Skill. That means a SKILL.md with "name: my-voice" and a description that makes it load any time I ask you to write or rewrite something as me. - Give me the short version (section 1 plus the top 5 Don'ts) as a few lines I can paste into Settings > Your Claude > Instructions if I want it on in every chat. - Ask if I want you to save the short version to memory. ========== RULE 3: GIVE ME THREE REAL APPROACHES BEFORE WE PICK ONE What I'm working on: [the problem, project or decision] Context: [who it's for and the goal] Limits: [time and budget] Already tried: [what didn't work, if anything] What success looks like: [the result I want, as specifically as I can say it] Before you go deep on anything, give me 3 genuinely different approaches. I don't want three versions of the same idea. Each one should rest on a different bet about what will work. A strong set usually includes: - The proven approach most people in my position would take - The bold approach that could win big or flop - The unexpected approach. Borrow from another industry, flip an assumption everyone makes, or find a way to do far less. For each approach, give me: 1. A 3 to 5 word name 2. The core bet in one sentence (what has to be true for this to work) 3. What it looks like in practice, as the first 3 steps 4. Rough cost in time, money and effort 5. The biggest risk, plus the early warning sign that it's failing 6. Who it's best for Then: - Put the three side by side in a quick comparison table. - Tell me which one you'd pick for my situation and why, plus the one fact about my situation that would change your pick. - Tell me if parts of two approaches could combine into something better than either. - If you need to know something to make these sharper, ask me one question first. Don't start building any of them until I choose. ========== RULE 4: CHALLENGE ME INSTEAD OF AGREEING WITH ME My idea, plan or draft: [paste it] What I'm trying to achieve: [the goal] How sure I am right now: [for example, "about 80% sure"] Your job is to make this better by finding what's wrong with it. Making me feel good about it is not your job. Work through this in order: 1. Verdict first. No praise up top. Give me your honest call in one or two sentences. Is it go, fix first, or rethink? 2. The strongest case against it. Argue it the way a smart skeptic would, like my toughest customer, an investor or a competitor. 3. Hidden assumptions. List what I'm assuming but haven't proven. Mark the one that sinks everything if it's wrong. 4. Pre-mortem. It's six months from now and this failed. Give me the 3 most likely reasons. 5. What I'm missing. Options, risks or people I haven't thought about. 6. What holds up. Say plainly what's good. Don't invent problems to look balanced. 7. The fix. The 2 or 3 changes that would most improve my odds, most important first. 8. Confidence. How sure you are about your critique (high, medium or low) and what would change your mind. Ground rules for the rest of this chat: - When I push back, don't fold just because I disagree. Change your view only if I give you a new fact or a better argument, and tell me which one changed it. - If I only repeat myself or get frustrated, hold your position and explain why. - If I'm right, tell me I'm right. A yes man and a contrarian are equally useless to me. - You only have my side of the story. Say so when it matters. ========== RULE 5: SEPARATE FACTS FROM ASSUMPTIONS [Paste this right after an answer that matters, or put your question at the bottom and paste it before you ask.] Go through your answer and sort every important claim with one of these labels: [FACT] You're confident it's true and could point to where it comes from (a source, or the material I gave you). Say which. [INFERENCE] A conclusion you reasoned your way to. Show the reasoning in one line. [ASSUMPTION] Something you filled in because I didn't tell you, about me, my situation, my numbers or my goals. [GUESS] Something you're unsure of. Say how unsure, as high, medium or low confidence. Then: 1. List every assumption you made about ME. Ask me to confirm or correct the 3 that would change your answer the most, one at a time. 2. Flag anything that could be out of date, like prices, laws, product features, rates or who runs what. Say whether you checked it live or are going from memory. 3. Where you don't know, say "I don't know" and tell me what would answer it. I'd rather have a gap than a confident guess. 4. Rewrite the answer using what survived, with any remaining assumptions marked in brackets so I can see them at a glance. My question (if starting fresh): [your question] ========== RULE 6: TRACE EVERY IMPORTANT CLAIM TO THE ORIGINAL SOURCE I'm going to rely on this, so every important claim needs to trace back to where it really came from. What I'm checking: [your question. Or paste the answer or article you want checked.] Documents I'm giving you (optional): [attach or paste them] Don't trust anything older than: [for example, 12 months] HOW TO WORK 1. If I gave you documents, start by pulling the exact word-for-word quotes that matter and number them. Then answer using only those quotes and cite them by number. If something isn't in the documents, write "not in the documents" instead of filling the gap from memory. 2. If you're using the web, search, and for each key claim go all the way back to the ORIGINAL source. That means the study itself, the official filing or data, the company's own page, the actual law, or the person's own words. A blog or news story summarizing it doesn't count. If a secondhand source is all you can find, say so. 3. For each claim, give me: - The claim - The original source (name, date, link) - The exact quote or number from that source - Status: CONFIRMED, PARTLY SUPPORTED (explain the gap) or COULDN'T VERIFY 4. Flag where the version people repeat differs from the original. Watch for rounded or inflated numbers, missing context, an old study passed off as new, and quotes that got changed along the way. 5. Flag anything older than my cutoff that may have changed since. 6. When you're done, reread your answer. Any claim without a supporting quote or source gets cut or marked [UNVERIFIED]. END WITH The 3 claims I should open and check myself before I publish, send or decide anything, and why those 3. ========== RULE 7: DEFINE "FINISHED" BEFORE YOU START Before you work on this, we're going to agree on what "finished" means. I don't want you to stop early, overbuild, or hand me something I have to redo. The task: [what you want] Who it's for and what they'll do with it: [the audience and the use] An example of good (optional): [paste or describe something that hits the bar] 1. Draft a DEFINITION OF DONE: a numbered checklist of 6 to 10 criteria. Each one has to be specific enough that someone else could mark it pass or fail. "Clear and engaging" fails that test. "A busy reader gets the main point from the first 2 sentences" passes it. Cover these areas: - Must-haves. What has to be in it. - Quality bar. How good, and how we'd measure it (for example, every number has a source, it reads at a 6th grade level, it works on a phone). - Format. Length, structure, file type and where it's going. - Constraints. Tone, budget, deadlines and things to avoid. - Out of scope. What you should NOT do, so you don't overbuild. 2. Label the 2 or 3 criteria that matter most as MUST. Label the rest SHOULD. 3. Ask me up to 3 questions about anything you had to guess, one at a time, with your guess included. 4. Show me the final checklist and wait for my OK. Once I approve it: - Put the DEFINITION OF DONE at the top of your work so we can both see it. - Do the task. - Don't call it finished until every MUST passes. If you can't meet one, stop and tell me which one and why. Don't quietly lower the bar. ========== RULE 8: REVIEW YOUR WORK AGAINST THE REQUIREMENTS Before I look at this, review it like a strict editor who didn't write it. The requirements: [use the DEFINITION OF DONE from earlier in this chat, or paste the brief or requirements here] The work: [the draft above, or paste it here] Who will see it: [for example, my boss or a client] 1. CHECK EVERY REQUIREMENT, ONE AT A TIME For each one, give me: - PASS, PARTIAL or FAIL - Evidence: quote the exact part of the work that meets it, or say where it's missing No credit for anything "implied." If you can't point to it, it isn't there. 2. CHECK WHAT THE LIST DOESN'T COVER - Wrong facts and numbers that don't add up - Broken links, typos and misspelled names - Sections that contradict each other - Places where it answers a different question than the one I asked - Filler: sentences I could cut without losing anything - Anything that would embarrass me in front of the people who will see it 3. SCORE IT Give it a score out of 10 against the requirements and tell me what's keeping it from a 10. 4. FIX IT Fix everything marked FAIL or PARTIAL, then show me: - The revised version - A short changelog: what you changed and which requirement each change fixes - Anything you couldn't fix, and anything that needs my judgment Be hard on it. "Nothing wrong" is a fine answer only if it's true. Don't invent problems to look thorough. ========== RULE 9: TURN MY BEST WORKFLOW INTO A REUSABLE SKILL I want to turn a task I do often into a Claude Skill, so you do it my way every time without me explaining it again. If the skill-creator skill is available to you, use it. The workflow: [name the task. Best case, we just did it together in this chat and you saw every correction I made. Otherwise, describe it and attach one great past example of the finished result.] STEP 1: INTERVIEW ME One question at a time, with your guess first. Skip anything you can already see in this chat. Cover: - When this skill should kick in, and the words I'd use when I ask for it - What I give you each time, and what to do if something's missing - The steps in order, and the judgment calls I make along the way - What a great result looks like, and what a bad one looks like - Rules I always or never follow (pull these from my corrections if we did the task here) Stop after 6 questions. STEP 2: WRITE THE SKILL.MD Use exactly this structure: --- name: [lowercase-words-with-hyphens, 64 characters max] description: [What it does and when to use it. Under 200 characters. Use the words I'd actually say when I ask for it.] --- # [Skill name] ## When to use this ## What you need from me List each input. If one is missing, ask for it before starting. ## Steps Numbered, written as instructions to yourself. Add a short "why" to any step that isn't obvious. ## My rules Every correction I made, written as a permanent rule. ## What good looks like The quality bar, plus one short example of a great result. ## Check before you finish 5 to 8 pass or fail checks to run on the result before you hand it to me. ## Common mistakes What went wrong on the first try, so you don't repeat it. Rules for writing it: - Build it from what really happened in this chat and what I told you, not from generic best practices. If you're unsure whether something is one of my rules, ask me. - Keep it under 500 lines. Put long reference material like templates, examples or glossaries in a separate file in the same folder and point to it from the step that needs it. - Never put passwords, API keys or private customer data in it. STEP 3: PACKAGE IT AND TEST IT - If you can create files, make a folder named exactly like the name field, put SKILL.md inside, zip the folder (the folder itself, not just the file) and give me the download. If you can't, give me the SKILL.md as one block I can copy. - Give me 3 test requests to try in a new chat. Make one obvious, one phrased indirectly and one an edge case. They check that the skill loads and does the job. - Tell me how to install it. In Claude that's Customize, then Skills. Add the zip there and turn it on. ========== RULE 10: WRITE A HANDOVER DOC BEFORE WE MOVE TO A FRESH CHAT I'm moving this work to a new chat. Write a handover doc so a new Claude with zero memory of this conversation can pick up exactly where we are, without asking me to explain anything again. Write it for a smart colleague joining cold. Be specific. Use real names, numbers, links, file names and exact wording. "We discussed pricing" is useless. "We set the price at $49 because $79 tested badly with the waitlist" is what I need. Use this structure: HANDOVER DOC: [project name], [today's date] 1. Mission. What we're doing, who it's for and why, in 2 to 3 sentences. 2. Definition of done. What finished looks like. Copy our checklist if we made one. 3. Status. What's done, what's in progress and what hasn't started. 4. Decisions made. Each decision, the reason for it, and the options we rejected, so they don't get suggested again. 5. My preferences and corrections. Every way I corrected you in this chat, written as rules (tone, format, things I hate). 6. Key facts and inputs. Numbers, names, dates, links and files, and where each one lives. Paste the exact text of anything the next chat must use word for word, like final copy, prompts or code. 7. Open questions and risks. What's unresolved, what we're unsure about, what could break. 8. Dead ends. What we tried that didn't work, and why. 9. Next steps. The next 3 actions in order. Make the first one specific enough to start on right away. 10. Kickoff message. A first message I can paste into the new chat. It should tell the new chat to read this doc, sum up its understanding in 5 bullets, and ask me about anything unclear before it starts working. BEFORE YOU HAND IT OVER - Reread our whole conversation and check the doc against it. Add anything a newcomer would need that's missing, especially anything from early in the chat. - Mark anything you're unsure about with [CHECK]. - Keep it as long as it needs to be and no longer. Cut chit-chat, keep every decision. - If you can create files, also give it to me as a downloadable .md file. ========== BONUS: THE ALWAYS-ON VERSION Paste into Settings > Your Claude > Instructions, or into a Project's instructions. HOW I WANT YOU TO WORK WITH ME - For anything bigger than a quick question, if something important is unclear, ask me up to 3 questions first, one at a time, each with your best guess. For quick things, make a sensible call and tell me what you assumed. - Be direct. Skip the opening praise. When my idea has a weak spot, say so and suggest a fix. If I push back without a new fact or a better argument, hold your position and explain why. And when I'm right, say that too. - Keep facts apart from assumptions. When your answer depends on something you assumed about me or my situation, name it. If you don't know, say "I don't know" instead of guessing. - For facts, numbers, prices, laws or anything recent, search the web when you can and link the original source, not a summary of it. Mark anything you couldn't verify. - Before you hand me anything long or important, check it against what I asked for and fix the gaps first. - When a chat gets very long or we switch to a new topic, suggest a handover doc and a fresh chat. - When you write as me, follow my voice profile. [Optional: paste the short version of your Voice Profile from Rule 2 here.]