cloud / systems
*Building features from the sofa sounds futuristic. In reality, it involves a surprising amount of waiting, checking and saying: “No, not quite like that.”*
I recently started managing parts of this website almost entirely from my phone.
The basic idea sounds wonderfully simple:
> Open ChatGPT, describe what you want, let it modify the GitHub repository, wait for CI/CD and deploy.
And technically, that works.
But there is an important difference between “possible” and “effortless.”
Prompt → production? Not exactly.
A typical request might start with something like:
> Add a magnifying-glass feature to the website, create a pull request, monitor the tests, fix problems and deploy it when everything is green.
That sounds like one instruction.
In practice it can become:
Prompt → PR → failed check → fix → another check → visual test → another fix → merge → deploy → test on iPhone → discover something strange → repeat.
The phone isn't the limitation. Reality is.
ChatGPT is not ChatGPT Work
The normal ChatGPT + GitHub workflow is noticeably more cumbersome than using ChatGPT Work or Codex for autonomous development.
One important example: when I leave the conversation, ChatGPT cannot simply sit there continuously watching GitHub. Scheduled checks may only run periodically — in my setup, roughly once per hour.
So if a pull request becomes green five minutes after I put my phone away, it may sit there for quite a while.
This is not quite the autonomous developer fantasy yet.
But it may be cheaper
There is another side to it.
Working this way consumes far fewer OpenAI coding credits than continuously using Codex or ChatGPT Work.
The cost does not disappear, though. CI pipelines, builds and repeated test runs consume GitHub Actions minutes and resources instead.
So the economics shift rather than vanish.
From my experience so far, this approach still appears cheaper for many small website changes — especially when the task is mostly prompting, reviewing and occasionally correcting rather than performing long autonomous coding sessions. I haven't measured a rigorous cost comparison.
Patience is part of the toolchain
The biggest lesson has probably been this:
AI development still needs supervision.
You need to test features yourself. You need to notice when something technically works but feels terrible on an iPhone. And occasionally you need to explain the same requirement again, just slightly differently.
The funny part is that I can now do much of that while walking around with my phone.
The less funny part is that I sometimes spend that walk watching GitHub status checks.
Still, the direction is fascinating.
My website repository, CI/CD pipeline and AI assistant effectively form a small remote development environment that fits into my pocket.
Not quite:
“Hey ChatGPT, build my website.”
More like:
“Hey ChatGPT, build this feature… no, not that way… check GitHub again… fix the test… yes, merge that one.”
And strangely enough, that already feels like a glimpse of how software development may increasingly work.
If you want to try the same setup, I described the GitHub connection and workflow in Codex credits exhausted? How I kept developing my website with ChatGPT and GitHub.