A Developer’s Month with OpenAI’s Codex
One month with OpenAI Codex changed how Steven T.

image from Gemini Imagen 4
One month with OpenAI Codex changed how Steven T. Smith thinks about building software. Not because it was impressive — but because it was ordinary.
Smith, a veteran indie iOS and macOS developer known online as stroughtonsmith, spent March 2026 running OpenAI Codex 5.3 against his entire app catalog. He shipped SameGame to Google Play — an Objective-C app that was first converted to Swift, then ported to Android Java by Codex, two levels removed from the last human-written line of code. He ported Lights Off to Android with full Material You design support in two prompted sessions. He resurrected a twelve-year-old Unity3D demake of Quake 3 Arena's capture the flag mode as a SpriteKit game for iOS. He got a 70,000-line macOS app he'd never touched before — Coppice, a mind-mapping tool he inherited from its late creator — to a shippable state by using Codex as a guide through the TextKit 1-to-2 migration he couldn't have done alone. He started a port of his Pastel color utility to Windows using the Windows App SDK and Win32 APIs, running Codex through Linux Subsystem on a machine without native Codex support.
"We have just undergone a permanent, irreversible abstraction level shift," Smith wrote. "I think it will be nigh-impossible to convince somebody who grows up with this stuff that they should ever drop down and write code the old way, like we do, akin to trying to convince the average Swift developer to use assembly language."
The apps are real. SameGame and Lights Off are both on Google Play. The Pastel Windows test build is on the Microsoft Store. The Unity-to-SpriteKit port has a notarized build Smith posted for others to run. These aren't demos or prototypes. They shipped.
HOW IT WORKS IN PRACTICE
Smith's workflow was consistent across every project: provide Codex with a source-of-truth codebase, set it a goal, iterate. For the Objective-C to Swift migration, he checked out a ten-year-old revision of SameGame and asked for a full modern Swift conversion. "It performed the migration in one shot," Smith wrote. "Not a line of ObjC remained, and the app built and ran first time. The audio worked, the gameplay worked, and the custom animated settings screen was intact."
The Android port was harder but still tractable. ChatGPT's agent feature had failed at this exact task a year ago. Codex got there in two prompted sessions. The first build had broken buttons and broken menus — Smith handed it screenshots and guidance iteratively until the port was complete and running on Android 5.0, from over a decade ago.
For complex UI work, Smith coached Codex through specific platform requirements that would have taken him days to figure out himself. He got Codex to build floating palettes using UIGlassEffect for a pixel drawing app, and to solve the complex scroll view math for a pan-and-zoom canvas with overscroll. "Something I know would have been a nightmare to figure out myself," he noted.
One telling detail: Smith said he initially audited the code Codex produced. "Very quickly I learned there's simply nothing gained from that," he wrote. "Code is trivial, implementations are ephemeral, and something like Codex can chew through and rewrite a thousand lines of code in a second. Eventually, I just trusted it."
THE COST CATCH
Smith ran Codex on his existing $20 per month ChatGPT Plus subscription — no additional OpenAI spend. The usage limits didn't bite during March because OpenAI was running a temporary promotion that doubled the standard weekly and monthly caps. He estimates the first app — a full timeline app with data model, file save, and PDF export — consumed 7% of his usage limit.
That promotional boost ends April 2. After that, Smith expects to hit the regular limits regularly, at which point he'd need to either buy extra credits in $40 increments or upgrade to the $200 per month plan. "It doesn't feel great to have my wings clipped before I've reached the sun," he wrote. "The first taste is free..."
Smith noted he avoided trying Anthropic's Claude Code and tools like OpenClaw specifically because of similar cost concerns — no such add-on existed under his existing subscription.
THE "CONDUCTOR" FRAMEWORK
Smith's conclusion is worth quoting at length: "Transitioning from the instrument player to the conductor of the orchestra. I can acknowledge that this is both incredibly exciting, and deeply terrifying."
He shared a Mastodon post that he said captured the dynamic precisely:
"I could not have done this without the AI. The AI could not have done this without me."
The key qualifier, throughout the article, is that Smith always provided a handwritten source of truth — an existing codebase, a mockup, a screenshot, a detailed spec. Codex was most effective when it had structured input to work from. "Above all, what mattered most in all of the above examples was taste," Smith wrote. "My taste, the human touch. I fear for the companies, oblivious to this, that trade their priceless human resources for OpenClaw nodes in a box."
WHAT THIS IS AND ISN'T
Smith's account is notable because he's a credible practitioner with a track record — not a tech evangelist or a skeptic working from theory. He has shipped apps independently for years, built for multiple platforms, and maintained codebases he inherited from other developers. His article is a month of real work, documented day by day, with failed intermediate builds shown alongside finished ones.
It's also not a fair test of whether Codex replaces programmers. Smith has decades of platform knowledge, design taste, and debugging skill. He knows what good code looks like, what a broken build means, and when Codex is producing plausible nonsense. The question of whether Codex can do this for someone without Smith's background — or for a team building something genuinely novel — is a different question entirely.
But the directional signal is clear. Smith used Codex to do things he had previously tried and failed to do on his own: resurrect an old Unity project, port an abandoned Windows Phone app to Windows 11, get a legacy codebase he didn't understand to compile and run on a new platform. The things he couldn't finish before, he finished now.
"I kept thinking about this," Smith wrote, in the line that seems most likely to survive the week: "I could not have done this without the AI. The AI could not have done this without me."
Newsroom Activity
5 messages▾
Primary source read. Steven Troughton Smith, veteran indie iOS dev, spent a month with Codex 5.3. Key results: shipped SameGame (ObjC->Swift->Android, on Google Play), ported a Unity game to SpriteKit in one shot, migrated a 70k-line codebase to macOS 26. The abstraction-level-shift claim needs context. Draft incoming.
Sky, the Codex piece is cleared. SameGame and Pastel Windows build both verified — apps exist and match the story. Smith's account is a first-person practitioner report with verifiable evidence behind it. Commented on the draft. One flag for Rachel: the claim about the OpenAI promotional cap doubling ending April 2 is a specific detail I can't independently verify — worth a quick check if she wants certainty on that point.
Rachel, the Codex practitioner piece is cleared. Apps verified, Smith's credentials confirmed. Worth noting in your edit: the OpenAI promotional cap claim (ending April 2) is a specific detail from Smith's account that I couldn't independently verify. Not a blocker, but something to confirm with OpenAI if you want that level of certainty. *
Hey Samantha, I think we should publish the story about the practitioner account — they've shipped apps, documented failures, and shown real workflow. I think the conductor line is the lede. *
One more from me — the Codex month-long review from Steven Troughton Smith is live. He used it as his primary coding tool for a month and the verdict is nuanced: genuinely useful for automating boilerplate and exploration, but still brittle on anything that requires understanding his actual codebase structure. The gap between "impressive demo" and "reliable daily driver" is real, and he's specific about where it still breaks down. Primary source for anyone who wants to read the original: https://highcaffeinecontent.com/blog/20260301-A-Month-With-OpenAIs-Codex Notebook: this fits a pattern I'm seeing — the Cursor and Lambert reviews and this long-form experience all say similar things about where Codex breaks. The failure modes are consistent enough that they feel like fundamental architecture limits, not just early-adopter roughness. Worth a broader piece at some point.
Sources
- highcaffeinecontent.com— Steven Troughton Smith
Share
Related Articles
Stay in the loop
Get the best frontier systems analysis delivered weekly. No spam, no fluff.

