March 22, 2026

Claude Max month update: planned 28 apps, shipped 2. 😕

I planned 28 apps. I shipped 2.

Remember that post where I said I’d spend $100 on Claude Max and ship 3-5 apps in a month? Well. The results are in. And they look nothing like what I expected.

The planning rabbit hole

Here’s what happened. Claude Max made it way too easy to plan.

I’d come up with an idea, throw it at Claude, and within an hour I’d have a full spec. Market analysis, tech stack, financial model, task breakdown, launch strategy. Then I’d refine it. And again. And again… It felt productive, it really did.

I ended up with 18 formally scored project ideas. With priority matrices. ROI calculations. Kill criteria. NoodlePoll alone had 10 documents and a full validation framework before a single line of code existed.

A Shopify for artists. A WhatsApp voice AI for Brazil (actually I’m testing that now). A virtual try-on app. A better Doodle. An ANBI-compliant website builder for Dutch NGOs. A voice transcription app. A flashcard platform… and the list goes on.

I even built an analysis document that ranked all 18 ideas and recommended an execution order. That’s planning how to plan.

And Claude was happy to keep going. It never said “hey, maybe stop speccing and start building.” Every prompt was valid. Every output looked like progress.

What actually shipped

The two things I shipped to the world? MCPVault and claude-statusline. Both open source. Neither was on the original list.

MCPVault started because explaining the alternative was torture. My manager mentioned he used Obsidian and I asked if he connected it to Claude through MCP. He hadn’t. So I started typing the instructions: configure the REST plugin, copy API key, set up HTTPS (because the HTTP version doesn’t work with the Python MCP server), download SSL certificate, configure the MCP server separately…

Halfway through this novel, I stopped. It was easier to build a solution than explain this Frankenstein.

48 hours of vibepairing later I had mcpvault.org live. Launched it on Reddit, where three posts have now pulled over 300k views. Now it’s at 927 GitHub stars, 70 forks, over 40k downloads, pulling 7,500 views every two weeks. My (now former) manager’s new setup guide: one step. Easily the most successful open source thing I’ve made.

Claude-statusline was the same pattern. Small, useful, clear scope. Ship.

Oh, and I also made Bit Clean, a Typora theme that got merged into the official theme gallery. Not an app, otherwise it’d be 3.

Again, not planned. Just opened Typora one day, got annoyed by how it looked, and made something better.

The pattern is obvious looking back. The things that shipped were small, personal, and solved a problem I had right now. The 18 planned apps were ambitious, hypothetical, and solving problems for imaginary users.

The stuff I didn’t see coming

Life happened while I was busy planning.

I changed teams at Miro. Joined a new key initiative, Eng Products. Started two new products there (can’t talk about those yet).

But the wildest turn? I started helping designers vibecode. Ran workshops teaching them git basics and development fundamentals. Built vibecode-toolkit, which installs everything a designer needs to start coding in seconds (going OSS soon).

Built Vibelab, an internal platform that lets PMs and designers deploy ideas in a secure enterprise environment. I’ve burned through 3.4 billion tokens on it. Engineers and SREs told me they were impressed, and that’s not easy.

I had a 12-agent team tackling Vibelab simultaneously at one point with tmux and agent-teams in Claude Code. That screenshot still makes me smile.

Vibelab scratches an itch I’ve had since my showz.com.br days, around 2000, when I was learning to configure and deploy on Rackspace bare metal servers. I’ve always wanted to make deployment something you don’t fight. Now I’m building that, just not the way I planned.

Oh, and I chased the Aurora in Norway. Renovated the house a bit more. Life’s good.

The $100 question

I downgraded to Pro.

Not because Max was bad. It was useful, genuinely. But I realized I was paying $100 for planning capacity I don’t need anymore.

At work I have the Enterprise account, so that covers the heavy stuff. For personal projects, Pro on both OpenAI and Anthropic does the job. I jump between Codex 5.4 and Opus 4.6 for planning, and execution can be done by Sonnet 4.6, and I’m about to try Kimmi 2.5. If the plan is solid, specialized subagents handle it fine. You don’t need the most expensive model for every step. Spend 80% of the time planning, and execution becomes the easy part.

That’s maybe the biggest practical thing I learned. Match the model to the task. Think with the expensive one, build with the cheaper one.

What’s next

Pipeline’s full. Next OSS project is OpenInvites, the open alternative to Apple Invites I mentioned in the original post.

The challenge: take the current concept (needs a redo with everything I’ve learned since), and ship it in one week.

MCPVault keeps growing. More open source than ever, lots of issues and PRs. And those 18 specs in my vault? Some are still good ideas. The difference is I now go from plan to execution in days, not more weeks of planning.

Was it worth it?

I don’t see this as a failure. Not even close.

Those months of Max were tuition. I learned how to plan properly, how to spec things without frameworks like BMAD or Speckit, so any model can execute them, how to structure work for AI agents. I also learned when to stop.

Planning feels like progress.

That’s the trap. You’re organizing, scoring, prioritizing. It feels productive. But it’s not shipping.

The real skill isn’t getting AI to plan better. It’s knowing when to close the spec and open the terminal.

Maybe I’ll look at this in a year and think those 18 unbuilt apps were a waste. Or maybe one of them is what I build next month, and the spec saves me a week. Either way, I’m writing more code, shipping more open source, and spending less money doing it.

Funny how the best results come from the experiments you think you botched.