I Tried to One-Shot an Entire App

Introducing the first app of the 52 apps in 52 weeks challenge: Kanboard.

The Problem

In the issue where I introduced this challenge, I mentioned the first product would help me manage all 52 apps.

Kanboard is what I envisioned as the most minimal project management tool for juggling many projects. A list of projects on the side, each with their own kanban board containing "To Do", "In Progress", and "Done" columns. That's it.

The reasoning: when I work on side projects like Puzzle Time, I just dump bugs and feature ideas into an Apple note. Kanboard is the evolved version of this. I can prioritize tasks, move things to "In Progress", and track multiple projects without losing my mind.

(And yes, I added keyboard shortcuts for everything because I love shortcuts.)

The 63-Question Experiment

Before building, I saw this tweet from Thariq about getting Claude Code to plan an entire product upfront:

I was skeptical but tried it. Claude asked me 63 questions. I was so shocked at the thoroughness that I tweeted about it:

It generated a full spec covering every detail. Authentication, team invites, calendar view, due dates, the works. I was pumped, so I set Claude loose to build the entire thing.

It ran for over an hour.

Plot Twist: Nothing Worked

Creating tasks didn't work. Dragging tasks didn't work. No optimistic updates. The landing page was rough. Basically nothing functioned.

You might be thinking "well, it got you 80% of the way there." Actually, the opposite. I spent so much time re-prompting AI to fix broken pieces that it would've been faster to build from scratch, piece by piece.

What Actually Matters

Don't try to one-shot an app unless AI can verify its own work.

Without tests or some feedback mechanism, it's just playing slots. If you give it solid unit tests, E2E tests, or you're building something like a CLI where it can run commands and check output, the odds improve dramatically.

In my case, Claude had zero idea if the code worked.

AI can't replace taste.

AI executes instructions. It doesn't think from the perspective of your user. It doesn't know the onboarding sucks, or that using the app daily feels clunky.

Now that we can delegate the boring implementation work to AI, we actually need to spend more time thinking, not less. More time on architecture. More time on UX. More time actually using the thing to feel what's broken.

One-shotting kills that feedback loop.

When you wait hours and burn credits on a massive generation, you have no idea how the app feels until the end. In my case, I ended up with features I didn't even want, like team invites and calendar views, that I had to rip out. Build incrementally. Feel the product as you go.

The Real Challenge Starts Now

Kanboard works now. It's simple, fast, and does exactly what I need. But getting here required throwing out half of what Claude generated and rebuilding with intention.

The lesson isn't that AI tools are bad. The lesson is that AI amplifies whatever process you give it. One-shot a messy prompt, get a messy app. Build thoughtfully in small steps with feedback loops, and AI becomes incredibly powerful.

Week 1 is done. 51 to go.

Try Kanboard here: https://kanboard.io/

Watch the full breakdown: https://youtu.be/zz7SitgcJBI

Let's build.

Anthony