Build your own apps with Claude, even if you've never coded A developer published a four-prompt method for building working apps with Claude, aimed at people who have never coded. The approach centers on keeping the model's context window clean using /context, /clear and a short CLAUDE.md file, and on having Claude interview the user to write specific instructions rather than relying on vague prompts. The writeup also covers running three builds in parallel and steps to take before sharing an app with other users. When I start a session, I ask one question first: which app would you happily never open again? Everyone has an answer. The timesheet tool with nine clicks to log one hour. The spreadsheet someone emails you every Monday. People answer fast, and a few get a bit angry, which is a good sign. Then people build, from the first minute. I don't demo for an hour while the room watches. Each person picks their own annoying app and runs the first prompt themselves. My favourite moment is when someone's own app opens on their own laptop, and they start listing the next three things they want to build. I wrote this for my friends who don't code and keep asking how I build things with Claude. You'll learn four prompts that take an idea to a working app, why Claude gets worse in long chats and how to stop it, how to run three builds at once, and what to do before anyone else uses what you built. Every prompt is here, in fold-out boxes. In this article Picture a librarian who has read nearly every book ever written. They remember no one, not even you five minutes ago. They answer only from the papers on their desk. That's the model. The desk is its context: everything it can see right now. Your messages, the files Claude reads, and the results so far. Each time you press Enter, the app sends the whole desk again, from the top. If something isn't on the desk, the model guesses. Fig 1. Every Enter sends the whole desk again. By message 50, most of it is old pages and failed tries. A new session starts with only the files. Illustration, not to scale. So long chats get worse. Message 50 carries the 49 before it, including every failed try. Anthropic's engineers call this context rot : as the desk fills, the model gets worse at finding what's on it. You'll see it. After a long chat, Claude forgets what you agreed and starts making things up. Typing more carefully won't fix it. So the method keeps the desk clean for you, in four ways: Most of what lands on the desk is what Claude pulls in while it works: files it reads and the output of commands. Your own messages are a small slice. Three things help: /context shows what's using space on the desk. /clear empties the desk and starts over. It costs nothing. A file called CLAUDE.md in your folder is read at the start of every session. Keep it short: what the app is, how to run it, how to test it. The security kit in Part 6 includes a template. The second half of the problem is the sandwich. Write the steps to make a peanut butter sandwich and hand them to someone who follows every word. They'll try to spread peanut butter with the jar still closed. AI does exactly what you write and guesses the rest. So in this method, Claude asks you the questions and then writes the instructions itself. Go deeper: the same ask, vague and specific 🔍 Vague: "Write a reply to this customer email." Specific: "Write a reply to this customer email for someone who isn't technical. Say what we found, what we did, and the one thing they must do next. Use 3 short steps and less than 120 words. No jargon." With the first one, Claude guesses who reads it, what it says and how long it is. With the second, you can send the result as is. The interview in step 1 writes the specific version for you. Small, real problems make the best first builds. Through this article I'll use one: a dinner plan for the week, built from what's in my fridge. Setup takes about five minutes. Get a paid Claude plan Pro is $20 a month , install the desktop app, open the Code tab, and pick a new empty folder. Then type: "Build a tip calculator. Keep it simple. When you are done, start it for me." If it opens and works, you're ready. Talk instead of typing when you can. The app has a microphone button, and people give far more context out loud. More first builds that work well: Who will use it decides what you build. Just you: a saved prompt or a skill instructions Claude keeps, so it does the job your way every time . Your team or family: an app. Hands-off: connect it to the system the data lives in, later, once the rules work. Start with the smallest one that does the job. No idea yet? Map one thing you do every week on paper, and mark each step: delete it, a fixed rule does it, AI does it, or I keep it. Go deeper: how to map a task in 30 minutes, with a worked example 🔍 A made-up example: a weekly status note, 8 steps and 72 minutes. | | Step | Min | Mark | |---|---|---|---| | 1 | Filter my finished tasks for the week | 5 | A fixed rule does it. A saved filter. | | 2 | Export the list to a spreadsheet | 2 | A fixed rule does it. | | 3 | Read each task and pick the important ones | 15 | AI does it. | | 4 | Look through chats for work that isn't in a task | 10 | I keep it. Only I know what happened there. | | 5 | Write the note | 20 | AI does it. I edit the first draft. | | 6 | Put the note in the team template | 5 | A fixed rule does it. | | 7 | Copy the note into a shared list | 5 | Delete it. Nobody reads the list. | | 8 | Send the note and answer questions | 10 | I keep it. | Step 7 goes first, because deleting it costs nothing. Steps 3 and 5 are the build. The minutes are how you'll know your build worked. I want to build something with AI, but I don't have an idea yet. Help me find one. I'm not a developer. Always talk to me in Simplified Technical English: short sentences, one idea per sentence, plain everyday words. If you must use a technical word, explain it in a few words the first time. Don't build anything yet. 1. Ask me what takes my time at work or at home. Ask up to 5 questions in one round. Number each question and give your recommended answer, so I can say "1 yes, 2 yes, 3 no". 2. If I pasted a map of one of my tasks below each step, the minutes it takes and its mark: delete it, a fixed rule does it, AI does it, or I keep it , start from the steps I marked "a fixed rule does it" or "AI does it". 3. Suggest 5 ideas. For each: what it does in one line, what it saves me, and if it needs an app, a spreadsheet, a document or just a saved prompt. Put the smallest useful one first. 4. Let me pick one. Then write it as one line that starts with "My idea:" so I can paste it into Step 1. Then stop. My task map optional : paste your steps and marks, or delete this line "But I can't code." You don't need to. You never write a line. Your job is to know what you want, answer good questions, and check that you got it. Claude picks the tools and writes the code. | Step | Where | What you have at the end | |---|---|---| | 1. Interview | New folder, new session | A numbered list of every decision | | 2. Spec | Same session | SPEC.md: the problem, what you get, how you'll check it | | 3. Tickets | Same session | A tickets folder, one small piece per file, plus STATE.md | | 4. Build | New session, same folder | A working app, every ticket built and checked | Nothing gets built until step 4. That surprises people. But most of the work is deciding what to build. The building is the fast part. Anthropic's own best-practice guide recommends the same opening move: ask Claude to interview you, then write the spec to SPEC.md. You paste the prompt with one line at the top: My idea: a dinner plan for the week, from what's in my fridge. Claude asks about 21 questions, in rounds. Each one is numbered and comes with Claude's recommended answer. So you can answer a whole round in one message: "1 yes, 2 yes, 3 no, it's for two people." Claude asks about the work: who uses it, what goes in, what comes out. It picks the tools itself and tells you in one line. Spend the most time here. When a build goes wrong, the cause is usually a question nobody asked in step 1. php Q1 - Who eats from this plan: just you, or more people? - Two adults. Portions sized for two. Q2 - Which meals: dinners only, or all three? - Dinners only for now. Q3 - How you'll tell it what's in the fridge: type a list, or take a photo? - Type a short list. A photo can come later. Q4 - What counts as a good plan: use up what spoils first? - Yes. Fresh food first, cans and frozen food last. You reply: "1 yes, 2 yes, 3 yes, 4 yes, and add the protein for each meal." The next round builds on those answers. I want to build something with AI. My idea: one line about what you want to build I'm not a developer. Always talk to me in Simplified Technical English: short sentences, one idea per sentence, plain everyday words. If you must use a technical word, explain it in a few words the first time. Data rules. Public information is fine. That means a published page, like a calendar, price list or set of rules, or a free public API that needs no login and no paid key and only gives out public data, like weather or public holidays. For a page, save a copy to a file in this folder and build from that file. If a page won't load for you, tell me and I'll copy and paste it in. No logins or passwords. Anything private, like work, customer or other people's data, is made up but looks real: realistic names, numbers, dates and amounts, in the same shape as the real thing. Never send private data to an outside service. Where a real system would send or get data, use a file, so a real export can replace it later. If this folder has the security kit a .claude/skills folder with secure-coding in it , use the secure-coding skill whenever you write code. Do not build anything yet. Interview me relentlessly until we reach a shared understanding. Map this as a design tree: every decision branches into the decisions that hang off it. Work the tree in rounds. The frontier is every decision whose prerequisites are already settled: the questions you can ask now without guessing at answers you haven't heard yet. Ask the whole frontier in one round: number each question and give your recommended answer. Then wait for my answers before the next round. Format a round like so: ❓ Q1 -