Claude said we fixed the SaaS MVP bugs, but I don't fucking know

This is a follow-up to a post I wrote today about how flawed my SaaS app really was. At the time I wrote that, it really was flawed. I was only midway through fixing the critical items Claude and Cursor had flagged.

Instead of the four issues I expected, it turned into seven. I honestly don't fucking know yet what happened there, or how well the code underneath it is written. That's a real weakness right now, but it's solvable.

This post is mostly for my own notes. Part of the journey, and maybe useful for anyone who wants to see how serious building something alone gets once you're past the fun part.

Seven bugs, not four

For the last 15 hours, I was heads down on the bugs and critical errors Claude and Cursor found during an audit. The real mistake was that Shared Workspaces was nowhere near ready. Roles and permissions were unclear, on top of a handful of smaller issues. I went through all of them one by one.

Even though Claude can act incredibly stupid sometimes, it does remember how I like to work.

Instead of ramming everything through in one prompt, Claude proactively proposed breaking each fix down, ELI5 style, and tackling one item at a time. That's always been my method. I don't one-shot things I don't fully understand.

So I went through every critical error, followed Claude's suggestions, and reported back after each prompt got pushed and committed.

We landed, allegedly, on a clean build. What Claude Code did is in the screenshot.

Then it told me to go build the Mac app

Out of curiosity, I asked a follow-up question: "ok, so what's the next step, or what should we do?"

The first response was to switch to the Mac app. Claude also mentioned I could keep working on the web app, or do something else, but what stuck with me was being told I could go fuck around in Swift next. That flustered me a little.

How confident does an AI need to be to tell you, don't worry, it's fine, go build a whole second native app on top of code we just spent a day patching?

That's the moment I fucking realized:

A lot of inexperienced vibe coders will just follow whatever the AI tells them to do, and they'll end up trapped in a loop of broken code and more bugs than they started with.

Read also: I build software like a toddler learning to walk

If I'd taken that advice at face value, I'd be two weeks into a half-ass-built Mac app right now, with a web app whose fixes were never verified.

That's why I didn't go with the Mac app. Instead I'm continuing to battle-test the web app, and I'm firing up a second account with a shared workspace to check whether everything Claude claims it fixed was delivered.

The temptation to just follow the advice is real, but I've proven to myself that shortcutting the QA step isn't going to pan out. I will just end up in a loop of more fuckery.

What's next?

I'm confident Claude Code and Claude understand my app now, four months into what's still basically a honeymoon phase, and I do trust the execution most of the time. But I'm not comfortable going aggressive on marketing yet.

My original plan to finish the Mac app in September isn't happening, and I'm fine with that. From here, I need to test every single item, take notes on bugs, and keep a running list.

I'm also missing a few things I should have built already:

• A proper user journey, a more detailed map of what happens when someone clicks A and lands on B

• A full pass over every feature I've marked as finished on the TWO roadmap.

I think that's two weeks of work, maybe through the end of September. But now I know not to touch anything else in the meantime. Just work through this list and keep documenting it publicly.

Today was still a good day. After a full day of checking what's missing on both sides, I'm more confident in where things stand and what's working on the web app.

Worst case, if my confidence drops once I've tested Shared Workspaces properly over the next week, I can just quietly disable access until I'm at least 90% sure it's solid.