Building Four Web Apps with Claude Code

I've been spending a lot of time lately building web apps with Claude Code, and I'm at the point where I have four projects on the go simultaneously. Two of them are nearly functional. One is beautiful but nowhere near done. And one is still basically a napkin sketch. Somehow, all four are teaching me something different.

The four are Timeliner, Commonplace, Waymark, and Badge Buster (that last one is still in the "great idea, zero code" phase). Each one started with a simple enough premise. Each one has made me want to throw my laptop into the Ping River at some point.

Timeliner and Commonplace: Functional First

With Timeliner and Commonplace, the goal was always utility. I didn't care if they looked gorgeous — I wanted them to work. Timeliner is a timeline visualization tool, Commonplace is for organizing quotes and passages. Both are React apps, both hit Supabase on the backend, and both have gone through more rewrites than I'd like to admit.

The challenge with this approach is that "just make it functional" sounds simple until you're staring at a broken graph view at 11pm wondering if the issue is the data, the component, or something you did six sessions ago that Claude doesn't remember. And that's the thing about Claude Code — it has no memory between sessions. Every time you open a new conversation, you're starting cold. Early on I'd jump in and say something like "fix the dark mode issue in Timeliner" and get back a perfectly confident response that was fixing the wrong file entirely.

What actually helped was building a CLAUDE.md file for each project — basically a brief for the AI that explains what the app does, how it's structured, what patterns to follow, and what not to do. Once I had that, the sessions got dramatically better. It's less about giving Claude more instructions and more about giving it the right context upfront.

Prompting also got sharper over time. Vague prompts get vague results. "The UI feels off" is useless. "The sidebar padding is inconsistent across breakpoints and the active state on nav items isn't visible in dark mode" is something Claude can actually work with. The more specific the bug report, the cleaner the fix — and the less time you spend reading through a response that technically answers your question but doesn't solve your problem.

Waymark: The Opposite Problem

Waymark is where I flipped the brief entirely. I wanted something visually striking — not just functional, but genuinely beautiful. A travel/location marking app with a design-forward UI that actually feels good to use.

That created a completely different set of headaches. When you're chasing aesthetics, you're dealing with CSS specificity wars, animation timing, layout quirks across screen sizes, and the constant tension between "this looks cool" and "this is a nightmare to maintain." Claude Code is actually pretty good at generating visually interesting interfaces, but keeping a consistent design language across multiple components over multiple sessions? That takes real discipline on your end.

I found myself spending as much time writing prompts that preserved what was already good as I did asking for new things. "Rebuild the card component but keep the hover animation and color system we established" is a sentence I've typed a dozen times. Without a design spec or explicit notes, Claude will happily recreate something in a totally different style and then you're backtracking for an hour.

Badge Buster: The Concept Stage

Badge Buster is the one that exists purely in my head, and honestly it might stay there — not because it's a bad idea, but because it's technically impossible.

The concept is simple: one button that clears every notification badge on your phone. All those little red dots on your apps, gone. I find them genuinely stressful and the only way to get rid of them is to open every single app one by one. Badge Buster would just... handle it.

Except it can't, because iOS and Android don't let one app reach into another app's notifications. Each app controls its own badge. So the whole premise falls apart at the OS level, which is a pretty fundamental problem for an app to have.

It's still on the list though, mostly because I think there's something there — even if the actual version looks totally different from the original idea. Maybe it's a focus mode thing, maybe it's something else. Sometimes a concept that doesn't work is still worth holding onto until you figure out what it's actually trying to solve.

The Stuff Nobody Tells You

Git was one of those things I kept putting off and then regretted every time something broke. Claude Code will make changes confidently, and sometimes those changes will break something else, and if you haven't committed recently you're doing archaeology trying to figure out what changed. I'm not doing anything fancy with GitHub — just committing regularly and writing halfway-decent commit messages so I can actually read the history.

The other thing is debugging with Claude. The instinct is to paste an error message and ask "what's wrong." That works sometimes. But describing what you expected to happen alongside what actually happened and what you already tried tends to get you to the answer faster. Treat it less like a search engine and more like a rubber duck that can also write code.

I'm nowhere near done with any of these four apps. Timeliner's UI still needs a proper library to replace the patchwork of manual styles I've accumulated. Commonplace has a few sync bugs I've been circling for weeks. Waymark needs a backend. And Badge Buster needs, well, everything.

But each one is teaching me to build better — clearer specs, tighter prompts, and the discipline to commit before you experiment. That's the thing about building with AI tools. The tool is only as good as how well you can describe what you need. Which, it turns out, is mostly just a writing problem.