Vibe Coding in Action Day 13: Building Multi-Platform AI Apps with Opus 5.5

An indie developer spent Day 13 of his live stream showing the unfiltered reality of building apps with AI coding tools day after day. There were no polished demos here — just a frontline developer grinding through usage limits, memory bottlenecks, and juggling multiple projects at once. Yet the stream inadvertently revealed the true capability boundaries of AI-assisted programming today, especially the "voice-controlled computer" feature he finally got working at the end — easily the highlight of the whole session.

The Reality of Managing Multiple Projects in Parallel

This developer maintains several products simultaneously: a web app, a mobile app, a video editing tool called Clipper, and an Input app for file conversion and screenshots. During the stream, he opened up about a significant technical migration decision — rewriting his native Swift mobile app and Clipper in Electron.

His reasoning was practical: while Swift works with his main cloud system, "there's no real connection between the internet and the cloud." Electron and cross-platform solutions are "more flexible." For something like Clipper, he admitted that "using Swift wasn't the best decision" — and even though the migration took a ton of work, he learned a lot from it. This kind of back-and-forth on tech stack choices is exactly what indie development really looks like.

The Usage Tug-of-War Between Opus 5.5 and Codex

A recurring anxiety throughout the stream was usage limits on AI coding tools. The developer kept switching between a $20 plan and an account with a 5-hour cap, watching the clock on every quota reset, and even trying to burn through remaining credits before his account expired.

"At first I was surprised by the usage on the $20 plan, but when you're working across this many projects at the same time, it starts to become a bit of a problem."

He had strong opinions comparing the two tools. Coming back to Opus 5.5 made a noticeable difference — he found it clearly stronger on both backend logic and design. His take on Codex, on the other hand, was pretty harsh: he mentioned Codex had "serious downtime issues" and even complained that even with daily quota resets, "the model is so bad I don't need those resets." He believes Opus 5.5 has "genuinely killed the case for a Codex subscription."

That said, he acknowledged his own usage habits were part of the problem: running 4 parallel agents at once and spinning up too many sub-agents caused his usage to spike and his Mac's memory to max out (24GB RAM completely filled, forcing him to use Codex just to clean up memory and kill RAM-heavy processes). "Too many sub-agents — it's bad for your usage" was a refrain he kept coming back to.

Opus 5.5 is Anthropic's top-tier model in the Claude 3.5 series (what the developer calls "Opus 5.5" likely refers to Claude Sonnet 3.5 or a newer internal version), well known for powerful code generation, long-context understanding, and complex reasoning — commonly used by developers for backend architecture and complex UI logic. Codex is OpenAI's earlier code-focused model, integrated into tools like GitHub Copilot, though its standalone competitiveness has faded with the rise of the GPT-4o series. Sub-agents refer to task execution units that a main agent automatically delegates to when handling complex work in AI coding frameworks like Claude's agent mode — each sub-agent consumes its own context and API call quota, so spinning up too many in parallel causes usage to grow exponentially, which is exactly why this developer kept hitting his limits.

Polishing Product Details: From Screenshot Selection to Auto-Updates

A large chunk of the stream was spent refining specific interaction details — and the way he described these requirements is itself worth studying. For the screenshot selection feature in the Input app, he wanted to remove drag functionality and instead have "the user click anywhere to start a new selection from that point" — simpler and better, he said, since "you can always undo anyway."

Other product improvements included:

• Unified file conversion entry point: Add a single conversion button for images and other files in the shelf, which opens a conversion interface with that file pre-loaded
• PDF generation: Support creating PDFs from text and images, with multi-file selection that layouts everything into a single PDF
• Thumbnail display: The file conversion interface should show actual image thumbnails rather than random filenames
• In-app auto-updates: Use a GitHub repository to implement in-app update checks and automatic download/install (combined with Cloudflare R2 storage and signing automation for Windows and macOS)

He also noticed a pleasant performance surprise: YouTube download speed in the Electron version was noticeably faster — "much faster than before."

The Big Highlight: Getting the Voice-Controlled Computer Agent Mode Working

Toward the end of the stream, the developer finally got his "Agent Mode" running — a computer-use agent that combines a local model for decision-making. He could give voice commands and the agent would automatically operate the computer: open Safari, type in a random email and password, search for keywords, check live Bitcoin and Ethereum prices, see what music was playing, and open a new YouTube mix that matched the vibe.

"We can actually control the computer with our voice now. You can literally talk to your computer and it will go do things. That's insane."

He was also honest about the current limitations: the agent can't click "accept terms of service and privacy policy" checkboxes, and it can't solve Cloudflare CAPTCHAs. This is exactly the bottleneck that all computer-use AI agents face today — they still struggle with anti-automation mechanisms and certain specific UI interactions.

His ideas for where to take this next were interesting too: give the agent a visible cursor so users can see what it's doing, let it pause mid-task to ask for input instead of just failing out, and even let it have a voice conversation. He thinks this has "enormous potential" for helping people learn new things. Getting this feature working took him about a week.

The "Computer Use Agent" is a frontier area in AI research. The core idea is to let the model perceive the screen through screenshots, then output mouse click and keyboard input commands to control a real desktop environment — Anthropic specifically demonstrated this capability when Claude 3.5 launched. Cloudflare CAPTCHAs (both Turnstile and traditional CAPTCHAs) remain a hard bottleneck for agents because they were specifically designed to distinguish humans from automated programs — they analyze the randomness of mouse trajectories, browser fingerprints, interaction timing, and other signals. AI agents operate too mechanically and are easily flagged as bots, making this a systemic obstacle shared by all computer-use agents today and one that's unlikely to be overcome through model capability alone in the near term.

What This Means for Indie Developers

The value of this stream isn't in any finished product — it's in the completely unfiltered look at what intensive indie development with AI assistance actually looks like: second-guessing tech stack decisions, carefully rationing usage costs, running out of hardware headroom, and the satisfaction of finally getting a key feature working after repeated trial and error.

A few real signals worth watching: cross-platform solutions like Electron still hold advantages in flexibility and certain performance scenarios; usage limits on AI coding tools have become a genuine pain point for heavy users; and while computer-use agents can already handle a fair number of real tasks, they're still getting tripped up by "small hurdles" like CAPTCHAs and terms-of-service checkboxes. All of these are key data points for gauging whether AI-assisted programming has truly matured.