AI development workflows — what can you do from your phone?

The other TDD: notes from the bathtub

The first time I built a feature from a bathtub, I laughed out loud. The water was still hot. Claude Code had just submitted a pull request, and I was viewing the results on a Netlify preview branch - the whole feature created from my phone. I knocked out a dozen or more feature and content changes in the same bath session.

I'm jokingly calling it Tub-Driven Development — the other TDD. The bathtub is mostly the joke, but the underlying shift it represents is real. Over the last couple of months, I've built websites, web apps and migrated a Drupal site to a modern Astro stack, almost entirely from my phone. Some of it from couches. Some of it while standing in lines. Yes, some of it from the bathtub.

I want to be careful here. None of this is to say software development is suddenly easy, or that the work of building good websites and web apps no longer requires solid technical knowledge. What's changed is where the work can happen, and which parts of it require sitting at a keyboard for it to be happening. The thinking is still the thinking. But the typing — the scaffolding, the boilerplate, the file-by-file plumbing — has loosened its grip on my typing time.

The setup

The stack is unglamorous, which is part of the point. Three pieces:

• GitHub holds the code. Nothing exotic.

• Claude Code has access to my repos through GitHub. 

• Netlify is wired to those repos for feature branch previews.

That's it. That basic setup enables a world of options on the go. 

The flow

Once it's wired up, a feature ships like this: I tell Claude Code what I want to change, look at the changes, commit to a remote feature branch and I review the results on a preview build. I ask for changes in plain English, but heavily leaning on planning first.

One caveat worth naming: decoupled builds have a real edge in this workflow. No SQL database to wrangle from a phone, no local environment that needs to mirror production, no pushing database updates between environments. Database-backed monolithic stacks aren't a blocker — they just take notably more setup and more care around environments and migrations.

The proof

The serious test was a migration. I had a site that had been on Drupal for nine years. I wanted to move to something modern — Astro, Markdown-based, deployed flat. The kind of project that used to mean blocking off a weekend and disappearing into a code cave.

I won't walk through the methodology in detail, because the interesting part isn't the recipe. It's that the recipe ran. The content modeling, the redirect map, the template and routing work, the deploy pipeline — all of it shipped from the tub. The "hard" parts of the project were the same as they always are: deciding what to do and catching the edge cases and mistakes. The "tedious" parts compressed by an order of magnitude.

What this isn't

This isn't an "anyone can build a website now" pitch, and I'd push back hard on anyone reading it that way. I can do the work because I've spent 25 years writing code, setting up environments, handling deployments and learning what a good codebase smells like. Claude Code is a power tool. Power tools in the hands of someone who doesn't know what they're building produce confidently wrong output at speed, which is worse than slow correct output.

What's actually shifted is the ratio of where my time goes. Less time typing, more time deciding. Less time in setup, more time in review. Less time in syntax and naming, more time in judgment. The senior engineering brain — the one that knows what good architecture looks like, what a maintainable component is, when to refactor and when to leave it alone — is more valuable than ever, not less.

I still love coding at my battlestation, I’m not giving that up for the phone. Real dev work still benefits from a real monitor (or four) and a real keyboard. Though honestly, give me Claude Design in the mobile app and I might cross another item off that list of desktop-only tasks. The bathtub isn't a destination, it's just now a viable place to do something I love — build things.

What it means for Alloy

The constraint on getting work done is so often defining the issue and defining the solution — things that take stakeholders, collaboration and sign off. Only once we work through those things can we build features, and how we build features is evolving fast because of AI tooling. That means a renewed focus on driving to decisions and enabling cross-team collaboration is key to truly harnessing the opportunities of these AI shifts. The percentage of effort in doing versus defining is flipping on its head, and previous understandings on what’s possible need to be reevaluated.

For our clients, it means the question of whether something is a good idea and gets a faster answer, and it probably gets an interactive prototype. It means I can be more present in client conversations and still have a first-pass feature planned and queued up to review by the time I’m back at my desk.

For our team, every developer at Alloy now has access to Claude Code and we've been spending real energy on what it looks like to be excellent with these tools. Not just using them — being good at using them, which is a craft in itself.

And for me, on a personal level, it means the unit of "I'd love to build that, but I'll never have the time" has gotten a lot smaller. Which is good, because the water gets cold faster than you'd think.