Building a paid Mac app in 11 days using Caude—no code written

Meta description: A ZDNET editor turned 2,338 vintage icons into a Mac app using Claude Code. Here's how 502 vibe-coding directives, App Store prep, and three days of promo art produced Icon Gallery '97.

Tags: vibe coding, Claude Code, Mac app development, App Store, AI coding agents

At around noon on Sunday, Sept. 20, ZDNET editor David Gewirtz came up with an idea for a Mac app. One week later, on Sunday, Sept. 27, also around noon, he submitted the completed app to Apple. Four days after that, Apple accepted the tool, and it went live on the App Store. The app, called Icon Gallery '97: 30th Anniversary Edition, was built through vibe coding — and it took 502 individual directives rather than one magic prompt. This article walks through what actually happened, what the AI did well, where it struggled, and what the experience means for anyone considering a similar project.

From a 30-year-old file to a working app

The story starts three decades ago. When Gewirtz began his first software company, one of his earliest products was Icon Factory for HyperCard, which won several awards and supported the business for years. In 1997, after years of customer requests for colorful Finder desktop icons, his company released a three-volume set of 32×32 pixel color icons called Icon Gallery — a boxed product containing more than 2,000 icons, sold through retail, mail order distribution, and direct mail.

Icon Gallery was one of the last boxed software products he produced, and it eventually faded along with physical manufacturing and the rise of the web. Months ago, while decluttering his office, Gewirtz found the packaging for a couple of old software products, including Icon Gallery. He stuck the box on a shelf and forgot about it again. Then, while writing historical Apple stories for ZDNET, he dug through his server and found an old Macromedia Director file containing all 2,338 icons from 30 years ago. In the mid-90s, he had converted the Icon Gallery icons to Director format and sold the result to animators as an asset library.

On Sept. 2, he handed that ancient file to Claude with a simple prompt: separate the icons into their own folder. Claude asked only what format he wanted and whether white edge pixels should be transparent. According to the log, the process took four minutes and 30 steps, producing 2,000-plus PNG images organized into topic folders. The extracted collection then sat in a folder on his Desktop for 18 days.

How 502 directives built the app

On a Sunday afternoon, needing a graphic for something else, Gewirtz thought it would be cool to have an app to browse and resize the old icons. That moment gave birth to Icon Gallery '97. He worked in both Claude Desktop (for strategy) and Claude Code (for building). Within a few hours, he had a working Mac app that could browse the original library and create perfectly scaled, hard-edge bitmap icons from 32×32 pixels up to 1,024×1,024 pixels.

Claude recommended two courses of action: naming the product Icon Gallery '97: 30th Anniversary Edition to celebrate the hand-created retro icons, and adding a recoloring feature so designers could match brand or identity colors. The AI also suggested collections and favorites. A friend recommended an "Export to Slack" feature for Slack emojis, and Gewirtz's wife saw the same potential for dragging icons into texts. Once favorites and collections existed, iCloud syncing became necessary.

To be clear, this was not a one-command process. Gewirtz gave Claude Code 502 individual directives across the four days he actively interacted with it. These were not full-time coding days — he visited the doctor for his yearly physical, spent Friday with a contractor working on the house, and turned in three major feature articles for ZDNET.

Where the AI pushed back

Gewirtz approaches AI coding agents as working partners at the end of a chat transcript, an approach he used for years with human programmers. That framing proved apt, because Claude pushed back on several instructions.

When Gewirtz asked whether they should care about someone buying the app, doing a full icon export, and requesting a refund, Claude suggested not fighting refunds — matching his long-term publishing style. But the question sparked a more interesting discovery: Claude realized all 2,146 icons sat inside the app as loose PNG files, meaning anyone with the free version could right-click, choose Show Package Contents, and copy the entire collection. They added encryption and packaged the app with a single scrambled file that unscrambles images dynamically as requested. Before submission, they deliberately corrupted copies to prove the check would catch them, then ran the app's own decryption engine over every icon; the decrypted output matched the originals.

At another point, Gewirtz noticed the window scrolled incredibly slowly. After he persuaded Claude to run tests, it turned out the code compressed full-sized images for each preview, and another block rebuilt the entire window 482 times during each resize request. Claude also initially thought a drag-selection rectangle was unnecessary — until Gewirtz explained that Macs have supported drag-select since before 1997.

Claude was notably careful about provenance, concerned about who owned the icons and whether the app would pass App Store review. Gewirtz had to convince it that his company created the icons in the 1990s and that his current business was the legal descendant. Even so, Claude flagged icons Apple was likely to block, including a Caterpillar logo and a person's head resembling Charlie Brown. A set of gunshot icons — popular in 1997 but tone deaf today — were removed as well.

App Store learnings and the promo art problem

Gewirtz is not a complete App Store neophyte, but he might as well be: his last App Store apps were built in 2008, and his last payment from Apple for them arrived in 2012. He used Claude to walk through each submission screen, taking screenshots of every new screen and sending them to Claude before hitting confirm or next. He also got advice on setting up the App Store account, including that new businesses under $1 million in sales can apply as a small business, dropping commissions from 30% to 15%. Before submitting, he instructed Claude to launch adversarial agents to examine the app end to end; they found a few things that needed fixing.

The biggest surprise was the promo graphics. Apple allows up to ten 2880×1800 graphics displayed with the app in a listing. Creating these took three days around his other work — almost as long as creating the app itself. According to Gewirtz, Claude, Gemini, and ChatGPT were "terrible, horrible, unfortunate, ineffective, and overall completely incompetent" at creating App Store promo images.

What this means for would-be vibe coders

The timeline is the headline: idea at noon on Sept. 20, submission at noon on Sept. 27, live on the App Store four days later. But the details matter more than the speed. A 30-year-old file in a dead format was extracted in under five minutes. A working prototype appeared within hours. Yet productization — recoloring, favorites, collections, Slack export, iCloud sync, encryption, performance fixes, and App Store assets — consumed the bulk of the effort.

For readers considering a similar project, the practical takeaways are these: expect hundreds of directives, not one prompt; treat the AI as a partner that sometimes argues; test security assumptions, because the AI may spot exposure you missed; and budget serious time for App Store graphics, which current image models handle poorly. Git branching also proved valuable — Claude built a completely different visual design within an hour, and Gewirtz decided in 30 seconds it was horrid, reverting instantly.

The app is now live. Whether it sells is a separate question, but the experiment answers one: a solo editor with a full-time job can ship a Mac app in a week using AI coding agents — provided he is willing to give 502 directives and spend three days on promotional images.