How I built a tax intake tool in forty minutes with eleven prompts

Last tax season, I built our personal tax intake tool the hard way. Two evenings in Airtable, with Claude as a thinking partner, iterating until it worked. This year, I rebuilt it in Lovable in forty minutes using eleven plain-English prompts. It works better than version one. I cannot write code, and I did not write any. I ran through the whole build live at The Firm's AI in Practice Summit, and this is the version of the story worth putting in front of practice owners.

The takeaway from that build is that you could build one too. You can go from an idea for a firm workflow to a working app in a set of prompts now. What used to take a developer weeks now takes an afternoon. That changes what practice owners can build themselves without waiting for a developer.

The first prompt is where the value sits

The most valuable thing in your first prompt is the boundaries you set.

When I opened Lovable and wrote the first prompt for the intake tool, the temptation is to start with the form itself. Build me a client intake form. Ask about tax year, income, dependants, deductions. That prompt gets you a dumpster fire you will iterate on forever, because the AI has no idea whether the form is for the client, for your team, or for both. It will mash the admin experience and the client experience into one interface that leaks client data to whoever opens the link.

The first prompt instead sets the schema and the boundary. Two tables: clients, and intake sessions. Two separate user experiences: an admin view for the internal team, and a client view for the person filling in the form. Those two boundaries decide the entire architecture of the app. Every subsequent prompt operates inside them.

The thinking that goes into the first prompt is design work, not typing work. I used Claude as a thinking partner to define the schema and draft the prompts before I pasted any of them into Lovable. The actual difficulty lives there.

The build sequence

Eleven prompts, in order, produced a working app.

Prompt one, the foundation and the boundary. Prompt two, the multi-step intake form with auto-save every thirty seconds so the client can close the tab and come back. Prompt three, year selection and resume logic. Prompt four, the client detail page for staff. Prompt five, the two-field new-client modal, name and email, because asking staff to type a client's address just duplicates the work the client is about to do anyway.

Prompt six, three transactional emails bundled into one prompt so they share templates, tone and sender. Asking for them separately would have produced three inconsistent emails. Prompt seven, AI-generated conditional questions comparing this year's answers to last year's. Prompt eight, the submission flow that updates the client record so next year's pre-fill is accurate. Prompt nine, the admin submissions page where the firm actually does the work of reviewing what came in. Prompt ten, real-time AI document verification. Prompt eleven, visual polish.

The pattern underneath the whole sequence is the same. Prove the substance first. Style it last. The tool has to work before it has to look good.

AI features are prompts inside the app now

The two AI features in the tool are the ones that make it useful. Conditional follow-up questions, generated by comparing this year's answers to last year's. Document verification, checking whether the file the client uploaded matches the document type asked for.

Five years ago, both of those features were developer builds. Someone had to write the logic, integrate the model, ship the code. Now they are prompts inside the app. AI is building the AI features. The system prompt you write is doing the real work, and you iterate on it the same way you iterate on any other part of the tool.

Design AI with an off ramp

The document verification catches most mismatches. It flagged a T5 uploaded as a T4 in the demo, and told the client to check. AI is sometimes wrong, though. A tool that blocks the client when AI disagrees is worse than a tool with no AI at all, because the client cannot submit their return.

The design decision is the override button. When AI flags a document, the client sees the warning, and they see a button that lets them insist they are right and continue anyway. If AI is correct, the client checks the file, replaces it, and moves on. If AI is wrong, the client overrides and submits.

The override is not a workaround. It is part of the design. AI in a client-facing tool has to leave the client a way to keep moving when it is wrong.

What is still missing

The version I built in forty minutes is not shippable. There is no authentication. Anyone with the URL could open the submissions and see the client's T4, their tax ID, all of it. There is no OAuth, so if your firm runs on Google or Microsoft, none of that is wired in yet. Error handling is light. Styling could be better.

Lovable has a security scan that runs when you try to publish, and it will flag the missing authentication. Run that scan. Add OAuth. Hold off on publishing anything with client data until security is properly in place.

Vibe coding gets you a working tool. Shipping it takes another set of steps that need to be done properly.

What to do on Monday

Pick one painful process in your firm. Ask three questions before you touch any tool.

What does success look like. Be specific. "I want a better intake form" is a vague goal that will produce a vague tool. "A returning client does not have to re-enter information we already have" is a real goal.

What does the data model look like. You already know this. You know what information you capture from clients, what you need to request, and what your team needs to see. That is your data model, even if you have never called it that.

Where could AI make a consistent judgement your team currently makes manually. For me, that was comparing this year's answers to last year's, and verifying an uploaded file matches the document requested. For your firm, it will be something else. Find the one your team keeps doing manually, and let AI do it consistently instead.

You can watch the full session here.