Guided build · an evening
Homework helper
Paste the problem; you and your kid write the solution steps together — and save them by subject to review before tests.
No commands. No code. You're about to build this by SAYING what you want — in your own language. The only skill is describing the result well. Here's the whole thing, start to finish.
Paste the problem; you and your kid write the solution steps together — and save them by subject to review before tests.
First, how to actually talk to the agent
- 1Describe the RESULT, not the tech: what's on screen, what each button does, what gets saved. The agent handles the 'how'.
- 2Be concrete about screens and fields — 'a form with amount, category and date' beats 'let me add expenses'.
- 3Ask for the smallest working version first. You grow it by talking afterwards — that's Step 2, not the first message.
- 4Decide what's YOURS, not the agent's: name the categories, colours, rules. Don't make it guess what only you know.
Step 1 — Build it live
Prompting for this kind of tool
- 1Pin down what one entry holds and how it's built: a pasted problem, a subject you pick (maths, science, English), and a solution written as numbered steps you can add to and reorder one step per line.
- 2Be explicit about the golden rule: the hint button NUDGES toward the next step, it does NOT hand over the answer. Say this clearly or you'll build a cheat machine instead of a learning tool.
- 3Spell out the saved side: saving stores the problem with its steps under its subject, and a saved list filters by subject so you can review before a test. Decide the subjects yourself.
- 4Always end with the magic line: 'build the whole thing and tell me exactly how to open it; don't stop to ask questions.'
claude --dangerously-skip-permissions
Open the terminal and run this (if Claude is already running, type /exit first). Then ask it to make a folder for this project and work inside it.
✕ what most people type
you ▸ do my kid's homework
✓ what actually works — copy it, or better, say it in YOUR words
you ▸ Build me a homework notebook for working through problems with
my kid. I paste in a school problem and pick a subject (maths,
science, English); below it we write the solution out step by
step, one numbered step per line that I can add to and reorder.
A 'hint' button nudges us toward the next step without handing
over the answer. Saving stores the problem with its steps under
its subject, and a saved list lets me filter by subject so we
can review before a test. Simple, friendly, saves on my device.
Build the whole thing and tell me exactly how to open it; don't
stop to ask questions.Say it in YOUR language — Russian, Spanish, anything. Claude understands.
Build it now — tick as you go
0/6Step 2 — It's built. Now make it yours
Don't panic at the techy link. Claude tells you one of two things. Either a file like index.html — find it in your project folder and double-click it; it opens in your browser. Or a web address starting with http://localhost — copy it into your browser's address bar and press Enter. 'localhost' just means 'this is running on your own computer right now'. That's your tool. It's real.
Now tweak it until YOU like it — just talk
This is the part that makes it YOURS. Don't accept the first version — change it by saying what you want, right there in the same chat. Claude rebuilds and you refresh. Keep going until it fits how you actually work. For this tool, try things like:
you ▸ add a 'hint' button that nudges us toward the next step instead of giving the answer you ▸ let me filter saved problems by subject and topic you ▸ add a quiz mode that hides the steps and asks my kid to redo a saved problem
I don't type long messages. I open voice-to-text (the mic in ChatGPT, or your phone keyboard), and just TALK through everything I think about what it built — what's wrong, what I want next — then select-all, copy, and paste the whole brain-dump into Claude Code. Messy spoken thoughts work fine; the agent sorts them out. It's several times faster than writing.
Step 2.5 — Test it like a real thing (it won't be perfect first try)
The truth no one tells beginners: the first build is a DRAFT. The agent makes something, you run it, you find what's off, you tell it — and THAT loop is the actual work. Don't read the code. Use the tool like a stranger would and watch where it lets you down. For this tool, check exactly this:
- 1Open it (Claude gives a localhost link or an index file — see Step 2), paste in a real homework problem and pick a subject.
- 2Write three solution steps, then drag to reorder them: do they actually add one per line and stay in the new order?
- 3Press the hint button: does it nudge you toward the next step WITHOUT just printing the final answer? If it spills the answer, that's your first fix.
- 4Save it, then open the saved list and filter by that subject: does your problem show up with all its steps under the right subject?
- 5Refresh the page (F5 / Cmd-R). Is the saved problem still there? If it vanishes, 'saves on my device' isn't wired up — fix that next.
Found something wrong? Say it like this
You fix nothing yourself — you DESCRIBE what's wrong in plain words and the agent fixes it. Be specific: what you saw vs. what you expected. Real examples for this tool:
you ▸ 'the hint button just gives the whole answer — make it only nudge toward the next single step and never reveal the final answer' you ▸ 'I can't reorder the steps — let me drag steps up and down and keep them numbered in the new order' you ▸ 'when I refresh the page my saved problems disappear — save them on my device under their subject so they stay'
Don't stop at v1 — grow it
A tool you built in ten minutes is a starting point, not the finish line. The people who get hooked keep saying 'now also…'. Each line below is one sentence you paste into the same chat to level this tool up:
you ▸ 'add a quiz mode that hides the steps and asks my kid to redo a saved problem from scratch' — review turns into real practice before a test. you ▸ 'let an AI check our finished steps and point out exactly where we went wrong, without giving the answer' — a patient tutor on the nights you're stuck too. you ▸ 'let me snap a photo of the textbook problem instead of typing it in' — kills the most annoying part, copying the question. you ▸ 'add a topic tag inside each subject so I can filter to just fractions or just verbs' — review gets surgical instead of scrolling everything.
Step 3 — Connect it to the world (optional)
Free data & APIs for this one
An API is just how your tool pulls in outside data or sends things out — many are free. You rarely need to research them yourself; the move that works for ANY idea is to ask the agent to find them:
- 1This tool gets much better with an AI service — Claude or OpenAI — to generate genuine next-step hints and check your solution without simply handing over the answer. Both use a paid key; a few cents per problem, spend under your control.
- 2Want to photograph a problem instead of typing it? Free OCR options like Tesseract (open source, runs locally) can read printed text from a photo — ask Claude to wire it in.
- 3Reusable move for ANY idea: ask Claude — 'look up free or low-cost APIs that help with tutoring and reading text from photos, and tell me which need a key.' It will list them; you pick.
About paying for the AI parts
- 1You don't pay for Claude twice. With a Claude subscription you create an API key right in the terminal and use it inside your projects — within your plan.
- 2From June 15 the rule splits: automation through that key gets its own monthly credit — about $20 on Pro, $100 or $200 on Max. Stay under it and it works as before.
- 3Run out and you pay per use (pay-as-you-go); unused credit doesn't roll over. So check your account and see whether the included credit already covers what you're building.
A tool gets superpowers when it can talk to other services — send you a message, read a sheet, pull live data. Each service hands out an 'API key' (a long password that says 'this is me'). For THIS tool, the useful ones are:
Which keys this tool can use
- 1An AI service (Claude or OpenAI) — give hints and check our solution steps without just handing over the answer
Getting any key is the same recipe: log into the service → find Settings → 'API' or 'Developers' → 'Create key' → copy it → paste it to Claude once ('here's my key: …'). Claude stores it and you never touch it again.
Step 4 — Put it online (only if someone else needs it)
Just for you, on your computer? You're already done — skip this step with a clear conscience. Put it online only if someone ELSE should open it, or you want to reach it from your phone away from home.
Step 6 — Stay safe (two easy habits)
- 1Back up anything you'd hate to lose. If the tool holds real data (your budget, health log, client list), ask Claude early: 'add an export/backup button' — one click saves your data to a file. Do it before you ever rebuild from scratch or move the folder.
- 2Don't hand Claude the keys to dangerous rooms. Keep it pointed at this project's own folder. Never connect a learning project to your bank, your crypto, or anything irreversible. Receipts and calendars — yes. Savings — no.
- 3Think before you ask. Claude does what you say in the folder you're in. 'Optimize my photos' optimizes them; it won't delete them — unless you ask for something reckless. Sane requests, clean workspace, and there's simply nothing scary left.
The mistakes that stop most people
- 1Asking for something vague ('make me an app') and getting something vague back. Name the screens, the buttons, what gets saved.
- 2Cramming in ten features at once. Ask for the tiniest working version first — then grow it with follow-up requests.
- 3Stopping at the first error. An error isn't failure — paste it to Claude or describe what you see, and it fixes it. That loop IS the work.
- 4Deciding you need to 'learn programming first'. You don't need to understand the code Claude writes. You need to describe the result.
- 5Forgetting the magic line — 'build the whole thing, don't stop to ask questions' — then wondering why it keeps pausing.
- 6Building for an imaginary audience. Build the version YOU will open tomorrow. Polish never; use immediately.
Normal. Tell Claude what's wrong in plain words ('the button does nothing', 'the page is blank') or paste the red error text. It finds and fixes it. You check. Repeat until it works.
That's it. You built a real tool.
Notice what just happened: no commands, no code — words. This is the entire skill. The people who 'can build things' are just people who describe what they want and let Claude do the typing.
More to build