Hey, it's Cindy 🌱 You commented apps, so here is the whole thing: the build manual you paste straight into GPT-6 Astra.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
the outcome
What you're making
One small app that does one thing, built on top of an image or video model you rent by the call. Not a startup. A product with one screen, one job, and a price on it.
The maths that makes this worth doing: at ten dollars a month, a thousand users is ten thousand dollars a month. You are not trying to beat Canva. You are trying to find one problem people already pay to solve.
step one
Find what's already earning
Two tools, both free to look at, no login needed to browse the charts:
📈 AppMagicappmagic.rocks/top-charts/apps. Pick a category, sort by revenue. The free plan also gives you competitor download and revenue estimates for last month, which is enough to check any idea.
🔎 App Store Trackerappstoretracker.com. Same job, different estimates. Worth checking both, because they rarely agree exactly.
⚠️ Two traps that will make you badly wrong. 1. Lifetime versus monthly. There is a table that gets copied around listing apps at "$13.2M" or "$27.7M". Those are cumulative since launch, not monthly. One of those apps actually earns about $300k a month. If you read lifetime as monthly you are out by roughly 40 times. 2. Net versus gross. Appfigures blog numbers are net, after Apple and Google take their 15 to 30%. Their downloadable reports are gross. Sensor Tower does not say either way. Never put two figures side by side without checking which is which.
the proof
What's actually earning right now
Every number below has a source you can open yourself. Read them as estimates unless marked otherwise.
🎬 AI Video - AI Video Generator (HUBX, Turkey). #16 in Photo & Video in the US, 4.5 stars from 27,000 ratings. AppMagic estimates over $1M a month. Look at what it charges: weekly subscriptions at $4.99, $6.99, $9.99, $12.99 and $14.99 a week, plus coin packs from $11.99 to $19.99. See the listing.
📸 Photo AI (Pieter Levels, indie). $105,000 a month revenue, $80,000 a month profit, published by the founder himself on 6 March 2026. It sells one thing: better photos of you, and it markets "dating photos for Tinder, Bumble and Hinge" right on the homepage. Read his post.
🧴 Cosmo AI (Spark Dynamic). A hair colour app. AppMagic estimated about $300,000 a month in August 2026 off 30,000 downloads. That is the whole product: change your hair in a photo.
✨ Remini (Bending Spoons). Plugged Google's Nano Banana model into an existing app in November 2025 and weekly installs jumped 175% to 1.16 million. Their announcement. Worth knowing the other half: Bending Spoons' own Q2 2026 filing reports Remini revenue declining. Trend apps spike and fade.
👔 Aragon AI (headshots). The founder gives the month by month on the record: $3,000 a month in February 2023, $180,000 by July, then $10M ARR in two years with 10 people on under $1M raised. The interview.
💡 The pattern: every one of these is a thin app around someone else's model, solving one problem, with a price attached. None of them invented a model. The offer is the product.
step two
Get your API key (and keep it safe)
🔐 Create the key Go to open.higgsfield.ai, open API keys, create a key, name it, and copy it. You get a key ID and a secret, two values, not one.
📁 Put it in a .env file, not in the chat Make a file called .env in your project folder and paste the values there. Then point Astra at that file. Your key never gets typed into a conversation, never ends up in a screenshot, and never gets committed to GitHub.
⚠️ Higgsfield's own rule: "Do not call the API directly from browser or mobile application code. Anyone who can inspect the application can extract its API secret and use your account." Your key lives on your server, never inside the app people download. Add .env to your .gitignore before your first commit.
step three
The build manual
This is the prompt from the reel, and it runs in GPT-6 Astra. Here is how to get it in front of Astra:
🗂️ Make an empty project folder Somewhere you will find it again, like Documents/my-ai-app. Put your .env file in it.
🤖 Open Codex and pick GPT-6 Astra Codex is where Astra can actually write files and run your app, in the ChatGPT desktop app or at chatgpt.com. Point it at that folder and select Astra as the model. Details in the Codex docs.
📸 Attach 2 or 3 screenshots Screenshots of apps whose layout you like. Astra reads them and builds to that layout, which beats describing a design in words.
🛠️ Paste the prompt and fill in your one problem Replace the bracketed line with the problem your app solves, then send it. Astra scaffolds the app, and you answer its questions as it goes.
🛠️ paste this into Codex
You're building me a working AI app on the Higgsfield API, and you're doing it in this folder. Read the docs at https://docs.higgsfield.ai/docs/llms.txt first, then build.
THE PROBLEM I'M SOLVING: [one sentence, e.g. "people need photos of themselves that don't look like selfies for dating profiles"]
Build the standard app interface for this kind of product:
1. An upload screen. One clear instruction, a drop zone, and an example of a good input versus a bad one.
2. An options screen. Between 4 and 6 preset styles as tappable cards with a thumbnail each. No free-text prompt box, presets convert better and fail less.
3. A progress screen. A real percentage from the polling loop plus a rotating status line, never a blank spinner.
4. A results screen. A grid of finished outputs, tap to enlarge, download all, and a share button.
5. A paywall after the first free result.
TECHNICAL REQUIREMENTS:
- Read HF_API_KEY_ID and HF_API_KEY_SECRET from .env. Never hardcode them, never expose them client-side. Add .env to .gitignore.
- Auth header format: Authorization: Key ${HF_API_KEY_ID}:${HF_API_KEY_SECRET}
- Generation is asynchronous. POST returns a request_id, then poll status_url until it reaches completed, failed, nsfw or canceled. Use backoff, not a tight loop.
- Before every generation, call the /estimate/ endpoint with the same parameters and log the cost in credits and USD. I need to know my cost per run before I set a price.
- Generated files are only retained for about 7 days. Download every finished output to my own storage immediately.
- Concurrency is the rate limit. Queue requests behind a worker pool rather than firing them all at once.
- Handle the failure states in the UI properly: failed and nsfw requests aren't charged, so offer a free retry.
Then run the app locally, show me the screens working, and tell me the cost per generation, the suggested price, and what my margin is at 1,000 users.
💡 Which model to ask for. One key gives you about 48 endpoints. Photo-style output, Higgsfield's soul models. Video from a still, kling, seedance or veo3.1. Fast image edits, nano-banana or flux-pro/kontext. Full list: the API spec. Swapping models is changing one string.
step four
Getting approved in the app stores
This is where AI apps actually die. Both stores have specific rules for apps that generate content, and the rejections are predictable.
Apple
📋 Guideline 5.1.2(i), the one that catches AI apps Apple: "You must clearly disclose where personal data will be shared with third parties, including with third-party AI, and obtain explicit permission before doing so." If a user's selfie goes to an AI API, you must say so plainly and get consent before you send it. Not buried in a privacy policy. On screen, before the upload.
🧱 Guideline 4.2, minimum functionality Apple: "If your app is not particularly useful, unique, or app-like, it doesn't belong on the App Store." A bare wrapper around an API is exactly what this rejects. Your presets, your onboarding, your saved history and your output gallery are what make it an app rather than a form.
💳 Guideline 3.1.1, payments Unlocking anything inside the app has to use Apple's in-app purchase. You cannot take payment on your own site to unlock in-app features.
📝 App Review notes Explain your AI feature in the review notes, including which third-party service you use. Apple clarified third-party AI disclosure in November 2025 and reviewers now ask.
Google Play
🛡️ You are responsible for what your app generates Google requires that generative AI apps do not produce content prohibited under its Inappropriate Content policies, and names the categories directly: non-consensual deepfake sexual material, voice or video that facilitates scams, deceptive election content, bullying and harassment, and apps primarily intended to be sexually gratifying.
🚨 Give people a way to report output A visible in-app report control on every generated result. This is the single cheapest thing to add and one of the most common reasons to get pulled.
🔍 Test your own prompts before you ship Try to make your app produce something awful. Whatever you can produce, a reviewer can too.
💡 Both stores, the same three fixes: disclose the third-party AI and get consent before uploading, add a report button to every generated result, and give the app enough of its own interface that it is not "just a wrapper". Those three cover most rejections in this category.
be honest with yourself
What this actually takes
🟢 One API call, today, for cents Key, pip install higgsfield-client, one subscribe() call. Half an hour, and you have a real generated result in your hand.
🟡 A working web app, this week Astra builds the screens above and you host it. This is where most people should stop and test whether anyone wants it, before spending anything.
🔴 Shipping to the app stores, this month An Apple Developer account is $99 a year, Google Play is a one-off $25, and both involve review. This rung is real work. Nobody does it in an afternoon and you should be suspicious of anyone saying otherwise.
⚠️ Don't chase a trend. The AI yearbook apps made millions in a fortnight in 2023 and are gone. Pick a problem that will still be annoying in two years, then the same build keeps earning.