From Idea to Prompt: The AI App Builder Translation Guide

Step 1: name the audience, narrowly

Not “everyone.” Not “millennials.” Narrow enough that the first 100 users know each other. “Bouldering gym owners.” “Recovering addicts using daily check-ins.” “Real estate agents in NYC.” The narrower the audience, the more the AI can specialize copy, features, and tone.

A narrow audience is also your best launch channel — you can find them in one community instead of 20.

Step 2: pick the wedge — one job

Your v1 app does one job exceptionally well. Listing all 12 features you imagine is how you end up with a generic app. Ask: “What is the single moment this app exists for?”

• Journal app wedge: “5-minute mood check-in at bedtime.”
• Task app wedge: “Capture anything in under 2 seconds.”
• Social app wedge: “Daily prompt everyone in the group answers.”
• Fitness app wedge: “Log one set without thinking about it.”

When the wedge is crisp, everything downstream (screens, data, tone) becomes decisive.

Step 3: list the screens — max 5 or 6

One line per screen. Name the screen + its primary job. If you can’t describe a screen’s job in one sentence, it probably does not belong in v1.

Good:
- Home: today's streak + one-tap check-in
- History: calendar grid, tap day for detail
- Settings: reminder time, paywall

Bad:
- Home (lots of stuff)
- Insights (advanced analytics)
- Social feed (friends' activity)
- Groups, challenges, leaderboards...

Step 4: name the data shape

AI builders produce generic mock data unless you give them a schema. Name the tables and the scoping rules. You don’t need to write SQL — one line per table is enough.

Data (Supabase + RLS):
- users (id, email, display_name)
- habits (id, user_id, name, emoji, reminder_time)
- check_ins (id, habit_id, user_id, date)

RLS: all tables scoped by user_id.
Offline cache: MMKV mirror of all tables.

When the AI knows the data, it stops generating fake “John Smith” avatars and produces realistic scaffolding.

Step 5: decide the visual tone

Not a brand document — four or five words of steer:

• Theme (dark / light / system)
• Accent color (hex)
• Density (airy / compact)
• Navigation type (tabs / stack / drawer)
• Typography mood if it matters (serif, monospace, rounded)

Everything more specific belongs in follow-up iterations, not the first prompt.

Worked example: vague idea → concrete prompt

Starting idea:

"An app that helps people stay in touch with
friends they never talk to."

After the five-step translation:

AUDIENCE: 28-40 year olds whose close friendships
slipped during the pandemic.

WEDGE: one daily nudge to reach out to ONE
specific friend — no feed, no scroll.

SCREENS (Expo Router, stack nav):
- Today: big card with one friend + prompt
  ("Text them now"), plus "skip" and "done".
- Friends: list, add from contacts, edit
  reminder cadence (weekly / monthly / quarterly).
- Insights: streak of days reached out, friends
  touched this month.
- Settings: reminder time, paywall.

DATA (Supabase + RLS):
- users (id, email, reminder_time)
- friends (id, user_id, name, phone, cadence,
  last_contacted)
- outreaches (id, friend_id, user_id, date)

TONE: warm, non-judgmental. Soft background
gradient, serif display typography, rounded
corners. Off-white theme, muted coral accent.

Same idea, but now the AI can build it. Before translation: generic social network. After: a specific, opinionated product.

Common translation mistakes

• Starting with tech (“I want to use Supabase and...”) instead of audience.
• Listing 10+ screens — guaranteed shallow output.
• Skipping the wedge — the app becomes “another X” instead of the distinctive one.
• Describing features instead of screens (“notifications” → which screen, which event, when?).
• Omitting data shape — you get mock arrays instead of realistic scaffolding.
• Over-specifying tone in the first prompt (“use Inter Display 800 at 48px...”) — save for iteration.

Use the framework as a checklist

Before hitting “Generate” in ShipNative (or any AI builder), verify you’ve answered: Audience? Wedge? Screens? Data? Tone? If any of the five is missing, your output will be 20% worse than it could be. The hour spent on translation saves a day of iteration.