AI Agents for Lawyers

Playbooks

Vibe-Code a Client Intake App as a Lawyer (Lovable, Claude Artifacts, v0)

Vibe-code a client intake app as a lawyer with Lovable, Claude Artifacts or v0: the spec, four exact prompts, a fee-band calculator and the IT handover.

By Dr. Niklas SchmidtUpdated 13 November 20268 min read
On this page
  1. What lawyers have built this way, and what it cost
  2. Choose the tool by where the data would go
  3. Write the spec before the first prompt
  4. The build, prompt by prompt
  5. Data handling: the rule that makes or breaks this
  6. Hosting, and what to hand to IT
  7. What to do this week

Vibe coding means describing an app in plain English and letting a model write, run and fix the code. A lawyer can build a working client intake app in an afternoon with Lovable, Claude Artifacts or v0, provided the prototype holds no client data, runs deterministic logic rather than a chatbot, and goes to IT before a real client ever types into it.

What lawyers have built this way, and what it cost

The most-cited example is not a vendor product. James Phoenix, a Linklaters associate, built an Excel plug-in in his evenings that cleans time-recording narratives; by September 2025 the "Narrative Checker", running on the firm's in-house chatbot Laila, was processing about 1,000 narratives a minute and had been used more than 300 times across 13 offices in its first week. Jamie Tso, a senior associate at Clifford Chance, has described building a track-changes analyser, a side-letter tabulator and a signature-page extractor in Google AI Studio for roughly nothing. A Texas family lawyer rebuilt a child-support calculator he had once sold on the App Store in under ten minutes with Gemini in June 2025, after Claude failed on the same attempt, a useful reminder that the tools are uneven and that trying two costs little. The community platform vibecode.law, run by Chris Bridges of Tacit Legal, Alex Baker and Matt Pollins of Lupl, reported in January 2026 that a hackathon had produced 20 working legal apps in a week.

The pattern is consistent: internal utilities that process the firm's own data, built by the person who had the pain. Deloitte's survey of 121 senior legal leaders, published on 30 June 2026, expects about 20% of in-house staff to become "hybrid engineer-lawyers", and Claude Code alone accounted for roughly 10% of public GitHub commits by March 2026, so the code is being written by models regardless of what the profession decides. A client intake app differs from a narrative cleaner in one respect that changes every design decision: a client touches it. The glossary entry defines the term; this page builds the thing.

Choose the tool by where the data would go

Tool (as of September 2026)What you getWhere it runsBest forData caution
Claude Artifacts (claude.ai, Pro from $20 a month)A single-page interactive app rendered next to the chat; code you can copy outYour browserPrototypes, calculators, forms that need not save anythingNothing persists, which is the point of a prototype
Google AI StudioTso's choice: free, builds and previews web appsGoogle's sandboxZero-cost first versionsConsumer terms; synthetic data only
LovableA full web app with a hosted database and a live URL from a promptLovable's cloudPermanent internal toolsA database means stored data: region, processor agreement and IT sign-off before anything real
BoltSimilar, with more visibility of the codeCloud, exportableWhen you want the code to leave with youSame as Lovable
v0 (Vercel)Polished front-end pages and componentsVercelA front-end a developer finishesFront-end only unless you add a backend
ReplitApp, hosting and a coding agent in oneReplit's cloudApps IT will adopt and maintainSame as Lovable

The view: prototype where nothing persists, then move. Claude Artifacts or AI Studio for the afternoon; Lovable, Bolt or Replit only once IT is in the room, because the moment the app has a database it has a data-protection footprint. Lawyers comfortable with a terminal can use a coding agent instead; the Claude Code page covers whether a non-programmer should.

Write the spec before the first prompt

Tso's four steps are the right order: pick one daily pain point; write the spec as a user's experience before asking for any code; choose the platform; iterate with screenshots. Skipping the second step is how a lawyer ends up with a chatbot when a form was wanted. The spec for this build:

Client intake app, spec v0.1

Users: prospective clients of a six-lawyer commercial firm, arriving from a link on the website; the intake assistant who receives the output.

Purpose: collect enough for a conflict pre-check and routing, and show an indicative fee band for three standard matter types.

Screens:
1. Welcome, privacy notice, consent tick-box (cannot proceed without it)
2. Who you are: name, company, email, phone, country
3. What it is about: matter type (pick one of three, or "other"), one-paragraph description, 500-character limit
4. Other parties: names of every other party and their advisers (required; add rows)
5. Urgency: any deadline, date picker; "none" allowed
6. Fee band: computed from matter type and two complexity inputs, from a table the lawyer maintains; labelled indicative, not a quote
7. Confirmation: summary on screen; "send to firm" button

Rules: no AI at runtime; every fee band comes from the lawyer's table; nothing is stored in the app; the summary is sent as structured text to the intake mailbox and the screen is cleared.

Out of scope: chat, document upload, payment, legal advice, storing anything.

Two design choices in that spec do most of the work. "No AI at runtime" means the app is a form with deterministic logic; a model writes the code, but no model runs when a client uses it. That removes the training and retention questions, avoids the EU AI Act's Article 50 duty for chatbots to identify themselves as AI (in force since 2 August 2026) and keeps the output predictable. "Nothing is stored" means the prototype cannot leak what it never held.

The build, prompt by prompt

Each prompt below is the whole message. Send it, test what comes back, and reply with a screenshot and one sentence about what is wrong; that loop is the method.

  1. The skeleton.
Build a single-page web app called "Intake" using plain HTML, CSS and JavaScript, no external services and no backend. It is a seven-step form for prospective clients of a commercial law firm. Steps: (1) welcome with a privacy notice placeholder and a consent checkbox that must be ticked to continue; (2) contact details: full name, company, email, phone, country; (3) matter: a dropdown with "Shareholders' agreement", "Commercial lease", "Employment termination", "Other", and a description box limited to 500 characters with a live counter; (4) other parties: a repeatable row with party name and adviser name, minimum one row; (5) urgency: an optional date picker for any deadline; (6) fee band, leave as a placeholder for now; (7) confirmation showing everything entered. Use a progress bar, Back and Next buttons, and keep the styling plain and professional. Do not store anything in local storage or send anything anywhere yet.
  1. Validation and the conflict fields. The other-party names are the point of the exercise; the conflict pre-check that follows depends on them.
Add validation: email format, phone with country code, and at least one other party with a non-empty name before step 4 can be completed. Add a field on step 4: "Is any other party a current or former client of the firm, as far as you know?" with Yes / No / Don't know. On step 3, if "Other" is chosen, require a short matter label. Show a clear inline error next to each field rather than a pop-up. Keep everything else unchanged.
  1. The fee-band logic, from a table the lawyer owns. The numbers are placeholders; the structure is what matters.
Matter typeComplexity inputsBand shown
Shareholders' agreementNumber of shareholders (up to 3; 4 to 6; more); bespoke vesting or leaver terms (yes or no)Three bands from a base figure; each step up multiplies by 1.5
Commercial leaseTerm (under 5 years; 5 to 10; over 10); tenant fit-out works (yes or no)Three bands
Employment terminationBoard member or executive (yes or no); pending dispute (yes or no)Three bands
Implement step 6. Put the fee-band table in a single JavaScript object at the top of the file so a non-programmer can edit it. For each matter type, two complexity questions decide a band: base, base x 1.5, or base x 2.25, shown as a range from 80% to 120% of the computed figure, rounded to the nearest 500. Base figures: shareholders' agreement 4000, commercial lease 3000, employment termination 2500; currency EUR. If matter type is "Other", show "We will confirm a fee estimate after a short call" instead of a number. Above the band show, in a visible box: "Indicative range only. Not a quote. A fee estimate is confirmed in writing in our engagement letter."
  1. Output and handover, with no storage.
On the confirmation step, add a "Send to firm" button that opens the user's email client with a mailto link to intake@example.com, subject "New enquiry: [matter type] - [company]", and a body containing all entries as labelled lines, other parties one per line, and the fee band shown. After the email client opens, show a thank-you screen and clear every field. Add a hidden second button, visible only when the URL contains ?test=1, that shows the same summary in a text box for copying. Confirm nothing is written to local storage, cookies or any server.

Expect three to six correction rounds: a Back button that loses entries, a counter that miscounts, a date picker that accepts the past. That is normal, and screenshots fix it faster than descriptions. The mailto handover is deliberately crude and deliberately safe: the enquiry lands in the firm's existing mailbox, which already has retention rules, and nowhere else. The n8n intake workflow shows what to do with it once it arrives.

Data handling: the rule that makes or breaks this

DataPrototype (Artifacts, AI Studio)Production (Lovable, Replit, own host)
Names of enquirer and other partiesSynthetic only, invented for testingReal, only after processor agreement, hosting region and IT sign-off
Facts of the matterNeverLimited free text; privacy notice; retention rule
StorageNone; session onlyYour jurisdiction; encrypted; access-controlled
AI at runtimeNoneNone, unless a reviewed business or API tier with no training and processing in your jurisdiction
Credentials and API keysNone in the projectIn a secrets store, never in the code or the chat
LogsNot applicableRetention set and documented

The reasons are not theoretical. The Upper Tribunal in UK v SSHD [2026] UKUT 81 (IAC) said that pasting client letters into public AI tools puts them "in the public domain"; the Austrian bar's four practice rules require client data to stay on EU servers and never to be used for training; BRAK's December 2024 guidance treats tools that store or forward client data for training as a breach of § 43a Abs. 2 BRAO. The Austrian and German guidance page has the detail, and the GDPR page covers the Art. 28 processor agreement any hosted version needs.

The coding tool itself is part of the risk. In July 2026 a wire-level analysis found that xAI's Grok Build command-line tool uploaded whole repositories, including git history and an unredacted .env file, to a cloud bucket: 5.10 GiB sent against roughly 192 KB the model used, and the "Improve the model" toggle changed nothing. xAI fixed it server-side on 13 July, and the same analysis found Claude Code and Codex sent no such bundles. The lesson is general: a coding assistant sees everything in the project folder, so the project folder must contain nothing you would not email to the vendor. Synthetic data throughout, and credentials nowhere near it.

Hosting, and what to hand to IT

A vibe-coded tool that lives on an associate's laptop is shadow AI with a nicer interface. The North Carolina Bar's January 2026 piece on realistic AI policies makes the point that bans drive shadow AI; a UK commentary on vibe coding inside legal teams recommends sandboxes with logging and lightweight approval rather than prohibition. Clio's 2025 Legal Trends Report found that 53% of legal professionals say their firm has no AI policy or do not know; if that is your firm, the policy template comes first.

What IT needs from you, in one folder:

  • The spec and the four prompts, so the app can be regenerated or audited.
  • The exported code in a repository the firm owns, not the vendor's workspace.
  • A data map: every field, where it goes, how long it is kept.
  • The hosting decision: region in your jurisdiction, an Art. 28 GDPR agreement with the host, and for German firms a § 43e BRAO contract.
  • Access control: who reads submissions, ideally behind the firm's single sign-on.
  • The privacy notice text for screen 1, approved by whoever owns privacy.
  • A test log on synthetic data, including an injection attempt typed into the description box; the agent security guide explains why that test matters even for a form.
  • A named owner and a kill switch: one person who can take the link down in a minute.

If all you need is the routing and not a client-facing app, a no-code stack may be cheaper: a Zapier and DocuSign intake setup documented in April 2026 costs about $210 a month and cut intake from five business days to same day. Module 5 of the AI Lab for Lawyers course builds an app like this in the browser, which is the right place to make the first mistakes.

What to do this week

  • Write the spec first, as a client would experience it, and show it to the intake assistant before any prompt.
  • Build the prototype in Claude Artifacts or Google AI Studio with invented names; four prompts, one afternoon.
  • Put the fee-band table in front of a partner and let them change the numbers; the app should survive that without a rebuild.
  • Book 30 minutes with IT and hand over the folder above; ask which hosting region and processor agreement they would accept.
  • Decide whether you need the app at all, or only the workflow behind it; the decision-tree playbook and the n8n build cover the two neighbours.

Frequently asked questions

What is vibe coding for lawyers?

Vibe coding is building software by describing what you want in plain language and letting an AI model write, run and correct the code while you test the result and give feedback, usually with screenshots. Lawyers use it for internal utilities: time-narrative cleaners, track-changes analysers, calculators and intake forms. You need no programming knowledge to start, but you do need a clear spec and firm rules on data.

Can I use Lovable or Claude Artifacts to build a law firm app?

Yes for prototypes and internal tools. Claude Artifacts renders a single-page app inside the chat with nothing saved, which suits a first version. Lovable, Bolt and Replit produce hosted apps with databases, which is where client data could end up, so use them only with synthetic data until IT has approved the hosting region, a processor agreement and access controls.

Is it safe to put client data in a vibe-coded app?

Not in the prototype, and in production only after IT review. Coding tools upload the project to their servers; in July 2026 Grok Build was found sending whole repositories including secret files. Use synthetic data throughout the build, keep the app a form with no AI at runtime, host in your jurisdiction under an Art. 28 GDPR agreement, and set a retention rule before a real client sees it.

How long does it take to vibe-code a client intake app?

A working prototype takes an afternoon: an hour on the spec, two to three hours across four prompts and screenshot-driven fixes. A Texas family lawyer rebuilt a child-support calculator in under ten minutes in 2025. Taking it to production, with hosting, privacy notice, access control and a test log, is a two-to-six-week job with IT, and most of that is not coding.

Do I need to tell clients the intake app uses AI?

If the app is a form with deterministic logic and no model at runtime, there is no AI interaction to disclose, which is one reason to build it that way. If you add a chat assistant, the EU AI Act's Article 50 transparency duty applies from 2 August 2026, so the assistant must identify itself as AI, and your bar's disclosure guidance and firm AI policy also apply.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion