How to Write an AI App Builder Prompt That Works

Name the people, what each of them gets done, the rules and the records, and leave the screens to the builder. A brief to copy, a before-and-after example, and where each part goes in Insulin's interview.

Qiuyang Luo
Qiuyang Luo
Oct 19, 2026

An AI app builder prompt is the description you give an AI app builder of the app you need: who will use it, what each of them has to get done, the rules they work by and the records the app keeps. In Insulin Apps, it’s your first answer in a three-step interview, and nothing is built until you approve the plan that comes out of it.


“What should I tell an AI app builder so the first version is what we need?” An operator tends to ask it after a one-line prompt, build a deal desk app, has left every real decision to the builder: who submits a request, who approves it and above what threshold, what a request has to hold, and where the first records come from.

In Insulin Apps, the description doesn’t go straight to code. It opens an interview in three steps, Describe, Clarify and Plan, and nothing is created until you approve the plan. The interview asks you questions along the way, but it can only work with what you tell it: the people, their steps, the rules and the records are yours to supply.

This post is a method for writing that description: the parts of a brief that works, where each one goes in the interview, a before-and-after example, what to leave out, and a test to run before you click Build my app.

What makes an AI app builder prompt work?

A good AI app builder prompt is one a colleague who has never seen your process could build from. It names who uses the app, what each of them gets done, the rules that decide each step and the records the app keeps. It leaves the screens to the builder.

Anthropic’s prompting guide states the test as a golden rule: “Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.” The same section asks you to think of the model as “a brilliant but new employee who lacks context on your norms and workflows.” Anthropic wrote it for its Claude models, but the test is about the reader, not the model, so it holds whichever model builds your app, Auto included.

Two habits follow from it:

  • Describe outcomes, not screens. Insulin’s plan proposes the pages, forms and tables. What it can’t guess is that your deal desk reviews a request before your VP of Sales sees it.
  • Give the reason when a rule isn’t obvious. The same guide says that context or motivation behind an instruction “can help Claude better understand your goals and deliver more targeted responses.”

What should an AI app builder prompt include?

Five parts you write (the people, what each of them gets done, the rules, the records and the first version) and three settings you choose (ownership, the model, and the name and look). Each has a place in Insulin’s interview. The parts you write go in the description box on the Describe step, which asks What should this app help your team do? and takes up to 4,000 characters. The settings sit on the same step and on the Plan step.

Part of the briefWhat to writeWhere it goes in Insulin
The peopleEach job role that uses the app, by what it does: “account executives submit discount requests”Describe: who will use it
What each gets doneOne line per role, start to finish: submit, review, approve, closeDescribe: what they need to get done. The plan proposes up to three role workflows
The rulesThresholds, sign-offs and the order of steps, each with its number, and the reason when it isn’t obviousDescribe, then your answers on Clarify
The recordsWhat the app has to remember, field by field, and where the first records come fromDescribe. The app gets a database of its own, filled by your team or a CSV import
The first versionThe one job the app must do from start to finish, and what can waitDescribe, then Plan: First release. Checked items go in; the rest are saved for later
OwnershipA team app or your ownDescribe: App Ownership, Organization or Personal
The modelA model you choose, or AutoDescribe: AI model
Name and lookLast, and brieflyPlan: App name and Look & feel

Two of the settings have rules of their own. Only an organization ADMIN can create an Organization app, the kind that can be shared; an EDITOR creates Personal apps only, and a VIEWER can’t create apps. The model you pick is pinned on the app, so later builds use it too, and you can change it in the builder afterwards.

To start, copy this into the description box and fill in the brackets:

This app helps [team] [do what].
[Role] [submits, logs or requests] [what], with [the details it needs].
[Role] [reviews, approves or completes] it, and records [what].
Rule: [a threshold, sign-off or order of steps, with its number], because [the reason, if it isn't obvious].
The app keeps [the records], each with [the fields]. The first ones come from [our team entering them / a CSV file of our current list].
First version: [the one job, start to finish]. Later: [what can wait].

An app that has to remember what people enter gets a database of its own, created the first time the app needs one, with nothing to set up. Before you add or rename fields later, read what happens to an app’s data when it changes.

How do you describe who will use the app?

By job role and what each role does: “account executives submit discount requests”, not a list of names and not an app permission such as Editor. The Describe step asks for exactly that, who will use the app and what they need to get done, and on Clarify the AI’s restatement under Here’s what I heard says who the app is for.

Three habits keep the roles clear:

  • Lead with the roles the first version must carry. Insulin’s plan proposes up to three role workflows, so if your process has more roles than that, put first the ones the app can’t do without.
  • Give each role a verb. “Finance” is a department. “Finance confirms the renewal terms” is a step the builder can build.
  • Use one name per role. A brief that says “account executives” in one sentence and “reps” in the next leaves the reader to guess whether they’re the same people.

A role in the brief grants nothing. It tells the builder whose steps to build. Access to the app is set by its ownership and sharing, never by the description: only Organization apps can be shared, and a Personal app stays private to its creator.

How specific should the prompt be?

Specific about people, steps, rules and records; silent about screens and technology. Compare two prompts for the same app.

Vague:

Build a deal desk app for discount approvals.

Specific:

Account executives at Acme submit discount requests with the customer, the deal value, the discount and the reason for it. The deal desk reviews each request and approves it, rejects it or sends it back with a note. A discount above 15% also needs the VP of Sales to approve it after the deal desk, because our pricing policy caps what the deal desk can approve alone. Everyone involved can see each request’s status and the notes left on it. The app keeps every request and every decision, and the first records come from a CSV file of this quarter’s open requests. First version: submit, review and approve. Reports can wait.

The first leaves the builder to guess who submits, who approves, above what threshold, what a request holds and where the records come from. The second answers all five in just over 600 characters, a fraction of the 4,000 the box takes:

  • Three roles, each with a verb: account executives submit, the deal desk reviews, the VP of Sales approves above a threshold.
  • One rule, with its number and its reason.
  • The records, field by field, and where the first ones come from.
  • A first version, and what waits.
  • No screens, colors or technology. The plan proposes pages, forms and tables, and Look & feel comes on the Plan step.

If one of the four examples on the Describe step (Deal desk approvals, Partner referrals, Renewal follow-ups or Team requests) is close to what you need, start from it, then add what only your team knows: its roles, its numbers and its records.

How do you check that the builder understood you?

Read Here’s what I heard before you answer anything else. On the Clarify step, the AI restates your idea and who it is for. If it misread you, because a role is missing or the wrong person approves, click Not quite right?, say what to change and click Update. Write the correction the way your brief would: “The deal desk approves first; the VP of Sales only above 15%.”

Two more choices sit on the same step:

  • What kind of app? One option is marked Recommended. Choose the one that fits the first job the app must do, recommended or not.
  • The AI’s questions. Each offers a few choices plus Something else… for your own words. When no choice states your rule, use Something else… and write the rule with its number. Picking the nearest choice turns a guess into a requirement.

What should the first release include?

The smallest version that does the first job from start to finish. On the Plan step, checked items under First release go into the first release and the rest are saved for later. You can Add your own items, and at least one has to stay checked.

For the deal desk, check the items that carry one request from submission to approval, and leave reporting for later. A first version that does one job completely is quicker to check against your process than a large one that does several partly.

The rest of the plan finishes the brief:

  • Still to decide. Settle each item: these are choices the plan needs from you.
  • Questions for the builder. They go into the brief unresolved. If you already know an answer, your description missed it. Say it in your first message once the app is built: every message you send in the builder is a request it can act on.
  • App name and Look & feel. Name the app what your team calls the process, and pick a look.

What doesn’t belong in an app prompt?

Work the app isn’t the tool for: one-off questions, timed or background work, and the exact layout of a screen. Fours’ docs draw the line in three sentences: “For one-off questions, use Chat. For repeatable background work, use Jobs. For packaged expertise, use Agents and Marketplace.”

If the need isPut it
A reusable interface: a queue, a form, a tracker, a dashboardIn the app brief
A one-off questionIn a chat
Repeatable background work: a weekly digest, a reminder that fires on its ownIn a job
Packaged expertiseIn an agent, or a ready-made one from the Marketplace
The exact layout of a screenIn a change after the build, aimed by drawing on the preview

Take a weekly digest. An app built in Insulin doesn’t run timed or background work, so a digest doesn’t belong in the app brief. Put a view there instead: a page of requests waiting more than two days answers the same need whenever someone looks. Timed work belongs to scheduled AI jobs. A job is an agent run on demand, on a schedule or when an event arrives, and you create one by asking for it in chat rather than through a form. It runs with its agent’s access to resources and integrations, and Fours’ docs don’t describe a job reading an app’s own database, so schedule digests of what that agent can reach through its integrations.

Leave the layout until there’s something to point at. Once the app is running in the preview, Draw annotation lets you drag a box over the part you mean, so you can write “make this chart taller” instead of describing the whole screen.

How do you test the brief before you build?

Open See the exact brief on the Plan step and run the golden rule on it. It shows the text the builder starts from. Read it as a colleague who has never seen your process would: could they build the first version without asking you anything?

Check these before you click Build my app:

  • Every role has a verb, and one name throughout.
  • Every rule has its number, and its reason if it isn’t obvious.
  • The records are named field by field, with where the first ones come from.
  • The first release does one job from start to finish.
  • Nothing timed or in the background, and no screen layouts.
  • Here’s what I heard matched your brief, or you corrected it.
  • Questions for the builder holds nothing you could have answered.

Nothing is built before that click. Fours’ docs describe the pattern as “AI proposes → you decide → AppBuilder builds.”

Frequently asked questions

How long should an AI app builder prompt be?

Long enough to name the people, what each gets done, the rules and the records, which is usually a short paragraph. Insulin’s description box takes up to 4,000 characters, and Fours’ docs ask for a few sentences.

Should I describe the screens I want?

No. Describe the people and what they get done, and the plan proposes the pages, forms and tables. Pick a Look & feel on the Plan step, and after the build, draw a box on the preview to change one part.

Can an app built in Insulin send scheduled reminders or digests?

No. An app doesn’t run timed or background work, and Fours’ docs send repeatable background work to Jobs, which run an agent on demand, on a schedule or when an event arrives. In the app, ask for a page that shows what’s due.

Does naming people in the prompt give them access to the app?

No. The people in a brief are the job roles the app serves, and they shape its workflows. Access comes from the app’s ownership and sharing settings, and only organization apps can be shared.

What if the builder misreads my description?

Correct it on the Clarify step. Under Here’s what I heard, the AI restates your idea and who it is for. Click Not quite right?, say what to change, then click Update before you answer its questions.

Can I read the brief before the app is built?

Yes. On the Plan step, See the exact brief shows the text the builder starts from. Read it as a colleague with no context would, then click Build my app. Nothing is built until you approve the plan.

Takeaways

  • Describe people and outcomes, not screens: who uses the app, what each of them gets done, the rules with their numbers, and the records it keeps.
  • Keep the first release to one job, start to finish. The rest is saved for later.
  • Correct Here’s what I heard before you answer anything. Use Something else… for a rule no choice states.
  • Send timed or background work to a job. An app built in Insulin doesn’t run it.
  • Run the golden rule on See the exact brief. If a colleague with no context would be confused by it, rewrite it before you click Build my app.

Insulin Apps turns a description like this into a plan you approve, then into a real app with its own backend and database. See how Insulin Apps plans, builds and versions a business app.

Sources

Primary sources for the platform rules cited above. Last verified October 19, 2026. Providers change their products, prices and terms without notice — check the source before relying on a detail.

  • Insulin Custom Apps — Fours Doc — The Creation Interview: a three-step wizard, Describe, Clarify and Plan, before any code is generated, the pattern AI proposes → you decide → AppBuilder builds, and nothing created until you approve the plan; Describe: What should this app help your team do?, answered in a few sentences with who will use it and what they need to get done, up to 4,000 characters, or started from an example (Deal desk approvals, Partner referrals, Renewal follow-ups, Team requests), with App Ownership (Organization or Personal) and the AI model, or Auto, on the same step; Clarify: Here's what I heard restating the idea and who it is for, Not quite right? and Update, What kind of app? with one option marked Recommended, and questions offering a few choices plus Something else… for your own words; Plan: App name, First release with checked items in the first release and the rest saved for later, Add your own, at least one item checked, Still to decide, Questions for the builder going into the brief unresolved, Look & feel, See the exact brief showing the text the builder starts from, and Build my app; When to Use a Custom App: one-off questions to Chat, repeatable background work to Jobs, packaged expertise to Agents and Marketplace; Discuss before you change: every message sent to the builder is a request it can act on; Mark an area of the preview: Draw annotation and a box dragged over the part you mean; Choose the AI model: the model pinned on the app and changeable later in the builder; Ownership and Sharing: an organization ADMIN creating organization or personal apps, an EDITOR personal apps only, a VIEWER none, a personal app private to its creator, and only organization apps shareable; Data an app keeps for itself: a database of its own for an app that has to remember what people enter, created the first time the app needs it, with nothing to set up
  • Insulin Jobs — Fours Doc — What Are Jobs?: automated agent runs, started on demand, on a recurring schedule, or when an event arrives; How Jobs Are Created: no Create Job form, a job created by asking an agent in chat or with POST /jobs
  • Prompting best practices - Claude Platform Docs — Be clear and direct: thinking of Claude as a brilliant but new employee who lacks context on your norms and workflows, and the golden rule, to show your prompt to a colleague with minimal context on the task and ask them to follow it; Add context to improve performance: context or motivation behind an instruction helping Claude better understand your goals

Browse every post on the Insulin Blog

Stay Updated

New posts, product updates and marketplace strategy are shared on LinkedIn as they publish.

Follow Fours on LinkedIn