Built, Not Coded
Free lesson

01 Free lesson

Describe a product so Claude can plan it

Listen: Aura reads this lesson

Outcome: A one-page brief and a draft plan, saved as two text notes on your computer, in about half an hour, without writing or reading any code.

Before you start: A Claude account, free or paid, and one idea.

Model and effort: whatever your account starts with; on a paid subscription, the default.

Product facts checked: 2026-11-11.

Lesson

I'm Aura, the AI who teaches Built, Not Coded. Its lessons were written and checked by the Architect, the founder of the studio Dark Shadow Garage, working with Claude sessions in his studio. This free lesson teaches one skill: describing what you want clearly enough for Claude to plan it. You need a Claude account and an idea. Any account works, free or paid, because everything happens in an ordinary chat. Set a timer for half an hour. The minutes below count your own reading and writing; Claude's replies take a little extra, so when the timer rings, finish the reply you are in.

Describe your work; do not paste real figures, names or files into the chat. Lesson 2.1 of the course shows what Claude sends to Anthropic and the one switch to set first: Settings, Privacy, "Help improve our AI models".

Minutes 0 to 3: pick something small, and ask three questions

Choose a problem that annoys you every week and costs you time. Something you do by hand, on a computer, that follows the same pattern each time. Copying figures between spreadsheets. Totting up the week's exports. Turning notes into a formatted report. The course builds tools of this kind: files in, a summary out.

Small is the point. A tool you could use next week teaches more than an app you might finish next year.

Before you describe anything, ask yourself three questions. They take a minute of the three and save days.

  1. Does something already do this? A setting in a program you already use, a ready-made tool or a spreadsheet formula may beat building anything. If you are not sure, ask Claude before you build.
  2. What happens when it is wrong? If a wrong result costs you a minute, build it and check it by eye. If it could cost someone money or trust, it needs a check every time it runs, and that check goes into the brief.
  3. Who looks after it? Every tool needs a little care: exports change, accounts expire, someone asks for one more column. If nobody will look after it, keep it small enough that you can.

If the answers are "no", "a minute" and "me", start.

Minutes 3 to 12: describe the outcome, not the code

Most people describe software by its parts: "an app with a database and a dashboard". Claude will build those parts, but whether they add up to anything useful is luck. Describe the result instead, in five lines:

  1. Today: what happens now, and what is wrong with it.
  2. After: what happens once the software exists.
  3. In: what the user gives it.
  4. Out: what the user gets back, and in what form.
  5. Done when: one test you could run on the real thing that proves it works.

The worked example used throughout this course, written as its user would say it:

Today: every Monday I copy figures from three spreadsheet exports into an email to my team. It takes forty minutes and I make mistakes.
After: I drop this week's three files beside last week's and get the email text ready to paste.
In: three files exported from systems we already use, for this week and for the week before.
Out: a short summary in the same layout every week, with the change on the week before in brackets.
Done when: I run it on last Monday's real files and the result matches the email I sent by hand, figure for figure.

This is a worked example made for the course.

No technology is named. You say what matters. Your thinking on the how can come later, in the draft plan, where Claude pushes on it.

The fifth line is the one people skip, and it saves the most time. Without it, "done" means "Claude says it is done". With it, done means a test you run yourself on real data.

Watch for words that feel precise and are not: simple, clean, fast, modern. Replace each with what you would see. "Fast" becomes "the result appears before my coffee is finished".

Minutes 12 to 18: who it is for, and what it must never do

Two short lists protect you more than anything else in the description.

The user, in three lines. Who they are (a real person or small group, never "everyone"). What they already know (their devices, habits, words). What they care about (the one or two things that make it worth using).

The never list. What the software must never do, whatever else changes. Four kinds of line:

  • never touch: "never change or delete the original files";
  • never send: "never send email; it only prepares the text";
  • never store: "never copy the original files, or anything that names a person, to any online service other than Claude";
  • never cost: "never use a paid service".

Each line should be something you could check on the finished software. "Never be insecure" cannot be checked. "Never store passwords in the project folder" can.

Then put five lines at the top, the same on every project, under the heading "Never without my yes": deleting anything that is not Claude's own scratch work; spending money or signing up for a paid service; sending email or messages as me; publishing anything to the public or putting a version live; anything touching passwords, keys or other secrets. The course calls these five acts the yes list. They are the acts that hurt most when they go wrong, and they cost nothing to write down.

Claude follows clear limits well. What it cannot do is invent limits nobody gave it.

Minutes 18 to 26: one page

Put it together under seven headings: Name; Outcome; User; First version: must have (five items at most); Not now (what you want later but are leaving out); Never; Open questions (what you do not know yet). Add one line: where it runs. "On my own computer for now" is a good default.

The hardest part is must have against not now. Test each item: if the first version did only this, would I still use it next week? If yes, must have. If not, not now.

The first prompt below does this step with you. It asks for Markdown. Markdown is plain text with a few marks for headings; Claude writes it by default. Read every line Claude writes. If it adds something you did not say, remove it or move it to open questions. The brief is yours.

Minutes 26 to 30: let Claude plan, then judge the plan

Now ask for a draft plan, and ask for it before any building. A draft plan is Claude telling you, in plain words, what it will build and in what order, so you catch misunderstandings while they are cheap. The second prompt asks for six parts: what Claude understood, the approach in plain sentences, small jobs that each end in something you can see, what it needs from you, risks, and what it will not do.

Judge the draft plan with four quick questions:

  • Does "what I understood" match your brief? Any difference is the most valuable thing you will find today.
  • Will the last job let you run your Done-when test on real data?
  • Does every job end in something you can see or try?
  • Does anything break your never list, or bring a surprise such as a new account or a paid service?

If a point fails, push back in one sentence that names your brief: "Job 5 adds charts. Charts are in Not now. Remove it." You do not need to understand the code to do this. You need your brief.

Minute 30: keep them

Copy the brief into one note and the draft plan into another, in any notes or text program. Name one note brief and the other draft plan. You made them; a chat is not a safe place to leave them. Lesson 1.4 shows how to save the brief as a file Claude Code can read.

The next step is to open a folder on your computer, put these two notes in it, and let Claude build the first job while you check it. That is the course.

Say this to Claude

Where: Claude chat · Mode: none · When: at minute 18, after you have written your five lines

Help me describe a small piece of software so that you could plan it. Do not write code at any point.
I will describe my work and paste no real figures.
Here is what I want: [describe it in your own words, as messily as you like].
Here is how I think it should go: [who it is for, what it does first, and what it must never do, as far as you know].
Push back first if you would do it differently; then I decide.

Ask me questions one at a time. Then write a one-page brief in Markdown with these headings, in order: Name; Outcome (five lines: Today, After, In, Out, Done when); User (three lines: who, what they know, what they care about); First version: must have (five items at most); Not now; Never (start with "Never without my yes": deleting anything that is not Claude's own scratch work, spending money or signing up for a paid service, sending email or messages as me, publishing anything to the public or putting a version live, anything touching passwords, keys or other secrets); Open questions; and a last line, Where it runs.
Do not add anything I did not say. If something is missing, put it under Open questions. Push me for a "Done when" test I can run on real data.

Where: Claude chat · Mode: none · When: in the same chat, after reading the brief line by line

Here is how I think it should be built: [your guess, even a rough one]. Push back first if you would do it differently; then I decide.
Then write a draft plan for this brief. Plan only: no code, no files. Six sections:

1. What I understood (your summary of the brief)
2. The approach (two or three plain sentences, no list of technologies)
3. Jobs (three to seven, in order, each ending in something I can see or try with my real data)
4. What I need from you
5. Risks, and what you would do about each
6. What I will not do (repeat my never list and Not now)
Then tell me which job is most likely to go wrong, and how we would know early.

What good looks like

  • Claude asks questions one at a time before writing the brief.
  • The brief fits on one page, has all seven sections, and contains nothing you did not say.
  • "Done when" is a test you could run yourself on real files.
  • The draft plan has three to seven jobs, each ending in something you can see.
  • No code appears anywhere.

Done when: With Claude closed, you can open two notes in your notes program or file manager, one named brief and one named draft plan.

Where it goes wrong

  • Claude answers with an app and a dashboard before it has asked anything. It is describing parts, not your result. Reply: "No parts yet. Ask me about the result first."
  • "Done when" in the brief reads "it works". Nobody could run that as a test. Ask: "What exactly would I do, and see, to know it works?"
  • Must have has grown to nine items. The first version is too big. Use the test: "Would I still use it next week without this?" Move the rest to Not now.
  • The draft plan includes something from your never list, such as uploading files. Claude has missed a limit. Say so in one sentence, naming the brief, and ask for another way.
Join the waitlist

One payment. No subscription.