Building a beautiful iOS app with 3 Claude Fable prompts

@anshuc
Anshu@anshuc
91 views Aug 02, 2026 ~7 min read
Advertisement

I worked on UI/UX at Apple for 12 years, so I've taken years off my life banging my head against UIKit, SwiftUI, and Xcode. I've used every major AI model for iOS dev, and they all sucked. Until Fable.

Media image

As a launch-day test, I asked Fable to built me a personal, delightful calorie tracker. I'm obsessive about calorie tracking and lifting, so I weigh and log everything. It's the worst. Every tracker app is ugly React/Flutter slop. Logging things takes a bunch of taps and searching. I wanted a simple, delightful app where I could just dump everything into an agent.

Fable built it with a single prompt (raw prompts at bottom of article):

0:28

This isn't just a mockup. It built the whole thing, including:

  • Nano Banana for image generation, with a custom Metal shader effect
  • Agent: Fable prompt engineered its personality, built its persistent memory, set up tool calling and token streaming, and even set up Anthropic prompt caching headers for it (never seen a model get this right before)
  • USDA database for accurate nutrition info
  • Nevertheless, it clearly didn't feel human-designed or polished: generic AI gradients, a messy layout, a few ugly pops and transition glitches. But with a couple more steps of asking it specifically to focus on aesthetics, we ended up with this:

    0:19

    Here are my learnings on how to get Fable to build amazing iOS apps.

    My tools and setup

  • I just used Claude Code in the terminal, with no fancy skills or plugins. Claude Fable 5 with xhigh effort. I think lower effort would have been better for human-in-the-loop design iteration, but I wanted this to be hands-off.
  • I'm on the 20x ($200) Anthropic plan. This used something like 1-1.5 M tokens across 4-5 iterations, over about 4 hours (though my involvement was <5 minutes). About 5% of my weekly quota. The $100 tier should work fine. The $20 tier can probably do this, but it'd use the whole quota.
  • A Mac with Xcode and the iOS SDK installed. No getting around that, unfortunately, if you want a truly native experience. But notably, Claude did not actually use Xcode. It just needs the Xcode tools (xcodebuild, xcrun) to drive everything from CLI.
  • I started in a completely empty folder
  • I keep a set of agent-specific, guardrailed API keys to common services (OpenAI, Anthropic, Gemini, etc) in a .env file that's sourced by my .zshrc, so terminal coding agents don't have to ask for these
  • Build a loop for Claude to check its own aesthetics

    Probably the highest leverage thing I did in my initial prompt was tell Claude to directly manipulate the simulator and test its work, and verify every interaction and transition works properly.

    Then during the second design pass, I specifically told it to verify every frame and pixel to the best of its ability.

    I didn't specify how, and didn't have to. Claude figured it out:

  • It brew installed idb to simulate mouse events inside the simulator
  • It used ffmpeg/ffprobe to record simulator usage and dump frames to examine transitions and animations
  • It wrote custom Python/PIL scripts to look at pixel differences between frames, crop and zoom frames, analyze pixel-level detail to fix pops, hitches, etc
  • The most important thing is setting the intent and expectation that you want a flawless, smooth, hitch-free interface, and that it needs to actually verify this. The model is smart enough to figure out the details, but you have to actually tell it to verify animations at the frame level.

    I didn't use /goal mode, but you could do so if you find it stopping prematurely. Fable seems to be trained to just pursue the goal non-stop anyway.

    The Prompts

    Original, one-shot prompt:

    I want you to build an iOS calorie tracker app. Should have a super clean, minimalist design, very design-forward and aesthetic. Sweat all the tiny details and make everything feel super premium. Make every interaction extremely delightful - you can go so far as to write custom Metal shaders, custom UI components, etc, to make everything fluid, unique, delightful.

    The idea I have here is to make it one unified conversational experience. I should be able to page through different days, each of which is a conversation thread with an agent. The agent should have whatever tools it needs to be able to query a nutrition DB or my past history and update the tracked food. It should feel personal and friendly.

    I want food entries to be beautiful - it should understand the foods and generate nice, consistent visuals for them, like studio photo looking images with consistent styling. You can use modern small/fast image generation models like Gemini or OpenAI; look it up. I want to be able to somehow zoom out of the thread view and be able to see all my meals in a scrollable timeline where each meal is represented using these images - should feel like a really premium magazine catalog kind of view. Make every transition fluid and delightful, fun gestures, bouncy, Apple-y liquid glass kind of feel to it. Use shader effects to make transitions look special and memorable, to make image generation look good, even to make the chat and token streaming experience unique and fun. Think through the flows to make it usable and show me the information I need without cluttering the screen or making it too technical.

    For the most part you should just use your best judgment, not ask me questions, unless you get hard stuck - I want you to interpret this however you want and show me what you are capable of. Build it end to end and verify all the interactions and transitions and functionality in the simulator if you can. Feel free to use the web to download any resources and documentation needed to make this feel great; don't just resort to the boring iOS standards for everything because that's easy. Dig deep and sweat every detail. Have fun!

    Start by initializing git, make commits as you go, keep yourself organized with docs so you don't lose track. Write clean, maintainable, modular code.

    Here's what I think was important about this:

  • I had a clear idea of the app I wanted, and explained the parts I actually cared about, while leaving the rest open-ended on purpose
  • I instructed it to be extremely meticulous about details, transitions, and "delight"
  • I asked it to verify all interactions and transitions itself, in the simulator
  • I gave it a rough workflow including documenting work so that it wouldn't lose track - I find this useful, especially across compactions, and for posterity if I end up with multiple agent threads on a long project. You want to document intent, not just the outcomes.
  • Follow-up to improve the design (this was a joke prompt I posted, but it did actually work pretty well):

    This is a great start but people on X are complaining it looks like AI slop because of the generic gradients, visual glitches, and awkward layout structure. So you need to cook hard and prove them wrong. I want you to 100x the design - get every pixel perfect, every transition flawless frame to frame.

    Find a way to carefully audit every frame/pixel and make it look absolutely stellar, like top 1% of human designers, instantly eligible for an Apple Design Award.

    Everything is on the table. Want to write your own components from scratch? Do it. Build a game engine for all I care. Write the most god forsaken Metal shaders the world has ever seen. Just make it look absurdly good. Don't stop until it looks like something Alan Dye would use. Make NO mistakes

    And some steering part-way:

    Consider a model that can do transparency properly. Make the high level grid view look nicer too; the alternating layout is a bit cheesy. Simplify, go more image-centric

    I want tiles to flow smoothly between the card and grid views. Build some gorgeous custom views/transitions, don't just use boring iOS view controller transitions.

    I don't think these were particularly good prompts but I'm sharing them unmodified for transparency's sake. I think all that's really happening here is:

  • I re-ran the model with its full attention on aesthetics; this time around, it didn't have to actually architect the functional app, so it was much more focused
  • Hyperbole about Apple design quality and going overboard might've helped a little
  • Specific steering is obviously very helpful to shortcut the time it takes to hone in on the worst visual issues
  • Telling it to aim for "pixel-perfect, every transition flawless frame-to-frame" is probably the most important instruction here, since that got it to actually record and dump frames and actually work until transitions were fluid.
  • In conclusion: The model goes very far with very little, but you sadly do still have to do the thinking. Give it ways to check its own work and you will be surprised by what it can do with a single prompt.

    Actions
    What You Can Do
    • Export as PDF or Markdown
    • Batch Export to Notion
    • Bookmark & Highlight
    • LinkedIn & Instagram Carousel Maker
    Create Free Account

    Includes 7-day Premium trial

    Advertisement