I see startups appear on X in the morning, where someone sticks together three APIs, a model, and a landing page, and suddenly there’s another “AI startup” with funding :)
But lately every engineer on the internet seems very excited to explain their “AI workflow”, so here’s mine too.
Somewhere around the end of last year I quietly entered that category of engineers who use AI constantly but still aren’t technically vibe coding. Not raw-dogging production code anymore (I do sometimes), but also not prompting way into a billion-dollar valuation.
It’s that where I still read the code, still debug things myself, but I would be lying if I said half of my commits didn’t start with an AI prompt.
Codex (my primary AI)
Most of my actual work happens inside Codex, the macOS app.
The UI is great and it feels properly native to macOS, which sounds like a small detail until you spend 8–10 hours inside a tool every day. Also yes, I am an Apple sheep so its much more obvious now.
Right now I'm mostly using 5.2. Apparently 5.4 dropped like six hours before I wrote this, so that will probably be the next thing I try.
Codex is basically where real development happens for me:
- writing code
- debugging
- exploring codebases
- refactoring things that past-me wrote
- explaining errors that somehow only happen on my machine
If something involves actual engineering work, it usually starts here.
Claude for thinking (the token hungry glutton)
When I need to plan something new — architecture design, system decisions, breaking down a new project — I usually switch to Claude.
Claude is still ridiculously good at structured thinking. If you want to talk through tradeoffs or design something cleanly, it does a great job.
The only downside is that Claude sometimes feels like a very enthusiastic way to burn through tokens.
Opencode for quick tasks
After adding NoeFabris/opencode-antigravity-auth Opencode basically became my AI scratchpad, whenever I need quick throwaway tasks, this is where they go.
Things like:
- spinning up a quick Redis Docker container
- generating command examples
- testing some config
- writing a tiny script
Nothing serious. Just the annoying little things that normally interrupt flow.
It’s basically the AI version of a sticky note.
Copilot, the OG tool
I’ve been using GitHub Copilot since 2022.
Which, in AI years, basically makes it ancient technology.
Inside my editor it still runs quietly in the background doing autocomplete. GitHub added Copilot Chat and a bunch of other things, but honestly I ignore most of that.
For me it’s just:
- good autocomplete
- less typing
- fewer boilerplate lines
It’s the quiet side character in the stack. Never dramatic, rarely impressive, but consistently useful. Which is funny, because that’s pretty much the exact role I seem to have in her story as well 😓
I also gave Cursor and Windsurf a fair shot.
They’re interesting products, but they feel a bit bloated for the way I work. And the credit systems never really made sense for my usage patterns.
The biggest thing I learned over the last year is that the best AI stack is usually the simplest one.
One tool for coding.
One tool for thinking.
One tool quietly autocompleting things in the background.
Everything else slowly turns into a messsss..
There are tons of tools in the ecosystem right now openclaw, ollama, and many more. They’re all cool projects. But they’re just too much for me.
