Content Creators Turn App Developers: How Vibe Coding Is Reshaping Software Production

From Reviewing Apps to Building Apps: A Spreading Creator Trend

In the tech content world, a curious shift is underway: the YouTubers and podcasters who spend their days reviewing, recommending, and complaining about apps are starting to build apps themselves. The Vergecast host David Pierce calls this phenomenon "creators becoming app developers"—especially in the productivity tool space, where many have spent years complaining that no to-do app fully meets their needs. Now, with tools like Codex and Claude Code, they're writing their own values directly into software.

This episode's guest, Stephen Robles, is a textbook example of the trend. He co-hosts podcasts including Mac Power Users and Primary Technology and is a deeply experienced player in the Apple Shortcuts automation space. Robles's path is highly representative: he first used AI to casually build a coffee-shop-finding app to test whether vibe coding could actually get an app onto the store, then gradually moved into Can Do, a productivity tool that people actually pay for.

Pierce's take on the show was blunt: this doesn't necessarily produce better software, but it does produce more software—and it changes how creators turn ideas into products.

Two Kinds of Apps, Two Kinds of Responsibility

Robles currently maintains two very different categories of applications, and he feels entirely different levels of responsibility for each.

One category is small single-purpose tools, such as Leveler (which automatically levels the volume of podcast recordings) and Captioner (which batch-adds animated captions to Reels). These tools often go from first prompt to launch in just two hours, and they're free and open source. Robles says he built Leveler because he was tired of manually adjusting audio levels every time he downloaded a video from Riverside, so he simply handed Claude both the "raw audio" and the "fixed audio" and asked it to rebuild an audio engine to match—it was working in twenty to thirty minutes.

The other category is Can Do—a new to-do/video project management tool with a kanban view, rich-text notes, and attachments. It started from Robles's love-hate relationship with Apple Reminders' kanban view: fast and native enough, but lacking rich-text notes and multi-attachment support. He built it for himself first, then released a TestFlight beta to the Shortcuts community, where the feedback was unexpectedly good, so he decided to keep investing.

The weight of responsibility is completely different for the two apps. If a free open-source tool breaks, users can just fork it; but Can Do carries users' expectations of migrating their entire workflows, which puts pressure on Robles as "a non-developer." He ultimately chose a $3 one-time purchase price—partly to recoup roughly $1,000 in token costs, and partly to deliberately set a barrier that prevents an explosion of feedback. He admits many people want him to switch to free-plus-in-app-purchases, but he "actually doesn't want that much feedback."

Claude Code in Practice: A Shippable App in a Weekend

Robles's AI coding workflow is already quite mature. His tool of choice is Claude Code (with the Fable 5.1 model) rather than Codex. His reasons are specific: Claude's generated UI matches his aesthetic better, so he doesn't have to repeatedly nag it to "leave more padding around the buttons"; and because he's been iterating in the same batch of conversations for a long time, Claude has accumulated a lot of context, so a small prompt can trigger a whole chain of correct changes.

More importantly, he handed his App Store Connect API keys and Cloudflare account to Claude. That means the entire release pipeline—from writing code and packaging the DMG, to uploading to the website, to generating App Store screenshots and filling in release notes—can be done in a single window. He described Claude automatically opening the simulator and taking screenshots on its own, and he thinks this kind of "plumbing work" is surprisingly well suited to AI.

His approach to bug reports is also highly personal: when users request features on social media, he uses a Shortcuts automation to collect screenshots into a desktop folder; when he sits down at his Mac, he dumps all the screenshots to Claude at once and says, "This is what people are complaining about—fix it." He also has a loyal user who sends a long bug list every week; he selects it all, pastes it, has Claude analyze which ones are fixable, and then says, "Alright, go do it."

This workflow isn't without its disasters. Robles mentions a "dark night of the soul": while trying to wire Siri AI-related features into Can Do, Claude repeatedly insisted the functionality was already built in, but when he tested it on his phone it simply didn't work—essentially being gaslit by the AI—and it was only resolved with an instruction file sent in by a listener. Complex integrations like iCloud sync and App Intents remain pain points for vibe coding.

vibe coding is a term coined by Andrej Karpathy in early 2025, referring to developers describing their intent to a large model in natural language and having the model generate or modify code, with humans mainly playing the role of "describing requirements—validating results" rather than reviewing the implementation line by line. What distinguishes it from traditional "AI-assisted programming" is the trust boundary: the latter treats AI as a completion tool while humans still control code structure; the former allows people to ship finished products without understanding SwiftUI details, so debugging tends to happen at the "behavior level" rather than the "syntax level." This also explains the phenomenon in the article where Robles says he "can't really read the SwiftUI code of these apps" yet still gets them onto the store—his abilities have shifted from writing code to breaking down requirements, organizing context, and validating results.

App Store Connect API is a REST interface Apple provides to developers, covering management actions such as app metadata, version builds, screenshot assets, review status, and sales reports. It requires JWT authentication using keys generated in the App Store Connect backend (Issuer ID, Key ID, and a .p8 private key). Handing such keys to an AI agent means the "administrative operations" in the release process can also be scripted: uploading builds, switching build versions, filling in release notes, and submitting for review can all be automated. The risk is equally direct—a leaked key equals a loss of account control, so in practice it's usually paired with least-privilege roles and separate CI credentials rather than long-term use of the primary account key.

Shortcuts, Siri AI, and Apps: Redrawing the Boundaries

As an authority in the Shortcuts space, Robles has a set of emerging judgments about "what should be an app, what should be a shortcut, and what should be handed to Siri AI."

His observation is that Siri AI has already swallowed a batch of shortcuts he used to rely on—for example, batch-parsing calendar events from screenshots, or recognizing YouTube videos in Instagram screenshots and jumping to them. But Shortcuts still has an irreplaceable place: when he produces Primary Tech Daily every day, what runs the RSS aggregation in the background is a shortcut; the "one-click copy of the latest podcast links across platforms" he made for Pierce is still a classic shortcut scenario—asking Claude to do that would just spin for five minutes.

The value of an app lies where a UI is needed. Adding animated captions to Reels involves fonts, positioning, and colors, and those must be carried by an interface. Robles's dividing logic thus becomes clear: things that require human visual decisions are apps; things that need to be stable, repeatable, and run in the background are shortcuts; one-off tasks with clear semantics go to Siri AI.

Pierce added a broader perspective: these three forms may not be converging—they may actually diverge further. Bespoke software handles "I do the same thing over and over"; shortcuts handle "assembling other software from building blocks"; and AI-generated-from-scratch suits one-off, exploratory tasks. Everyone has to re-decide which parts of their life belong in which category of tool.

App Intents is a framework Apple introduced starting with iOS 16 to expose an app's capabilities to the system in a structured way, so that actions can be invoked by Shortcuts, triggered by Siri voice, and appear in Spotlight search and system-level suggestions. Its key difference from the older SiriKit is that intents declare their own parameters and return types, with the system handling semantic matching and parameter filling, so the same set of definitions can serve both the voice and automation entry points. For indie developers, App Intents is almost the only proper way to "be seen by the system"; but its debugging chain spans processes and system services, and errors often manifest only as "I said it and nothing happened"—which is exactly why iCloud sync and App Intents are pain points for vibe coding in the article: the model struggles to infer the real fault point from error messages.

A Creator Flywheel, or Just More Software?

Robles doesn't call himself an "app developer"; he describes himself as "an app project manager and vibe coder"—he can't really read the SwiftUI code of these apps. He also isn't planning to give up Shortcuts as his main work; instead, he sees a flywheel: the apps he builds can include Shortcuts actions, he demonstrates those actions in his videos and podcasts, and that in turn drives usage of the apps.

His view of the future carries a straightforward optimism: in five to ten years, building your own software will become so easy that not just creators but ordinary people will start doing it. He's even started running local models with LM Studio to save tokens and building shortcuts that can call local models, believing this opens yet another door.

Pierce's summary is sharper: people who used to monetize their communities by selling courses would, in many ways, be making huge progress if they switched to building software. And app development becoming a core part of creator identity may be a change he didn't anticipate.

Whether this trend ultimately settles into a better software ecosystem or merely produces more long-tail tools, one thing is already certain: writing code is no longer the barrier. Deciding what to build, who it's for, and whether you're willing to take responsibility for the user experience—that's the new dividing line.

One-line Summary

Content creators are using AI coding tools to transform from app reviewers into app developers, reshaping software production and the boundaries of responsibility.

Paragraph Summary

Using podcaster Stephen Robles as a case study, the article describes a new trend of tech content creators building apps themselves with tools like Claude Code. Robles maintains two kinds of applications: free open-source utilities that can launch in two hours, and a paid to-do tool, Can Do, that must take responsibility for users' workflows. He hands App Store Connect and Cloudflare keys to Claude, enabling a full release pipeline from coding and packaging to screenshots and release notes, while admitting he was "gaslit" by AI on iCloud sync and App Intents. The article further maps the redrawing of boundaries among apps, Shortcuts, and Siri AI, and argues that writing code is no longer the barrier—deciding what to build, for whom, and whether to take responsibility for the user experience is the new dividing line.