AI Guides › Step-by-step guides
By Nigel Guy · 7 min read
Most people build a small tool by typing "make me a quote calculator" and then spending an hour steering the result with corrections. It feels productive because something appears on screen straight away. What you actually get is Claude's guess at your job, which you then patch one complaint at a time.
The rule: write the one-page Build Sheet first (who uses it, what goes in, what comes out, what it must refuse to do), then let Claude build from the sheet, one change at a time.
You need a Claude account and a job small enough to fit on one screen: a calculator, a checklist, a tracker, a converter, a small form that produces text. Claude's help centre describes artifacts as the place where Claude puts things like "a small interactive tool" beside the conversation, and says they work on Free, Pro, Max, Team and Enterprise plans.
Cost at time of writing: the Free plan is £0. Pro is listed at $20 a month (about £15 at time of writing; check the £ price at checkout, as VAT and exchange rates move it). Some artifact features, such as data storage, are listed for paid plans only, so a tool that must remember entries between visits is a Pro-or-above job. Plans change often, so check the current pricing page.
You do not need to install anything. If you later want a tool that lives as files on your own computer, Claude Code is the route. It is included in Pro and above according to Anthropic's pricing page, but it is a separate and heavier path, so this guide stays with the browser.
For the worked example, imagine a hypothetical window cleaner who wants a quote calculator. Swap in your own job as you go.
Before opening Claude, write these six lines in a note. This is the sheet.
| Line | What goes in it | Hypothetical example |
|---|---|---|
| Purpose | One sentence: who uses it, to do what | I use it on the doorstep to give a price in under a minute |
| Inputs | Every field, with type and allowed range | Number of windows (whole number, 1-60), storeys (1-3), conservatory yes/no |
| Rules | The arithmetic or logic, in plain words | £4 per window, plus 25% for the second storey and above, minimum charge £25 |
| Outputs | What appears and how it is formatted | Total in £, a line-by-line breakdown, a copy button |
| Refusals | What it must not do or must flag | Reject zero or negative windows; show a warning above 60 |
| Done when | Three checks you can run yourself | 10 windows, 1 storey = £40; 1 window = £25 (minimum applies); 0 windows shows an error |
The "Done when" line is the part people skip, and it is the part that lets you tell a working tool from a convincing-looking one. Work out the expected answers by hand before Claude builds anything.
Open a new chat at claude.ai. Paste the prompt below, with your sheet filled in. If you see an Output option in the message box, you can choose it, but a normal chat works as well: Claude creates an artifact automatically when the result is self-contained and substantial, and you can ask for one directly.
You are a careful front-end developer building a small single-page tool for a non-programmer. Build it as one self-contained artifact I can use immediately.
THE BUILD SHEET
Purpose: [ONE_SENTENCE_PURPOSE]
Inputs: [FIELDS_WITH_TYPES_AND_ALLOWED_RANGES]
Rules: [THE_LOGIC_IN_PLAIN_WORDS]
Outputs: [WHAT_APPEARS_AND_HOW_IT_IS_FORMATTED]
Refusals: [WHAT_IT_MUST_REJECT_OR_WARN_ABOUT]
Done when: [THREE_CHECKS_WITH_EXPECTED_ANSWERS]
What a good result looks like: it works on a phone screen, uses British English and £ formatting, has clear labels, and shows a plain-English error beside any invalid input rather than failing silently.
Order of work:
1. Read the sheet. If any line is missing, ambiguous or contradicts another, ask me up to five questions and stop. Do not guess.
2. Once I answer, build the tool.
3. Before you show it, run my "Done when" checks against your own logic and tell me for each one: expected answer, your tool's answer, pass or fail.
Constraints: no external services, no accounts, no data sent anywhere. Do not add features that are not on the sheet. Keep the layout simple. List any assumptions you made at the end in a short bulleted list.
Fill in the six bracketed lines from your sheet. Answer any questions Claude asks, since each one is a gap in your sheet that would otherwise have become a wrong guess.
Do not rely on Claude's self-report alone. Type your three "Done when" cases into the tool yourself. Then try the awkward ones: an empty field, a negative number, a very large number, letters where numbers belong.
Ask for a single change per message, and say what should stay the same:
Change only this: [ONE_CHANGE]. Leave everything else exactly as it is, including the rules and outputs on the Build Sheet. Afterwards, re-run my three "Done when" checks and report pass or fail for each. If this change conflicts with the sheet, tell me instead of making it.
Bundling five changes into one message is how a working calculator quietly breaks. If a change goes wrong, the artifact keeps earlier versions, and you can also edit an earlier message in the chat to branch a different version without losing the previous work, as the artifacts help page describes.
Each time a rule changes, change it on your sheet too. The sheet is the source of truth; the artifact is just its latest build. Without this, you will not remember in a month what the tool is supposed to do.
Artifacts start private to you. To share, open the artifact and click Share. On Free, Pro and Max, publishing makes it available to anyone with the link. On Team and Enterprise, sharing stays inside your organisation. Access levels include Can view, Commenter and Can edit, and email invitations (listed as beta, up to 50 people per artifact) are available on Pro and above. Check the current sharing page, since these details are still moving.