6-week challenge: we're not gonna make it

I recently set myself a challenge: build and launch a small product in six weeks using AI tools.

I didn’t get it done.

A big reason is because I didn’t give myself enough time over the six weeks. The second I started the challenge, work decided to get very busy. But, that’s also not the whole reason.

It was definitely a lot harder than I thought to build something fully usable using Replit.

In hindsight, I probably should have spent a little bit more time scoping, building, and deciding on the product before committing to the six weeks. I hit the wall I think a lot of people hit with AI coding tools — you get 80% of the way there very fast, it feels like you’ve almost built the product, but then that last 20% takes way longer than the first 80%.

Here’s what I was building

The product was called Proposal GPT.

The idea was to help consultants and agencies generate polished client proposals fast. You would chat with the tool, input rough notes about the proposal you’re creating, and it would use AI to generate a clean client proposal that you can send to a client.

I started breaking the project down into milestones:

• Let the user chat and add rough notes and generate a proposal → ✅ I got there

• Let users edit the proposal via chat → ✅ got there

• Let users co-edit the doc with AI in real time → ❌ didn’t get there

• Let users download a PDF → ❌ didn’t get there

• Let users create custom templates, save documents for later, select from multiple templates, etc. → ❌ got some of this working, but didn’t get it finished

• Add logins, team plans, make a proper SaaS → ❌ didn’t get there

I got the MVP kinda working. But once I started layering on anything more complex, the whole thing started to break.

Where it got tough

A few big issues came up for me:

1 - Scoping projects with AI tools is way different to how you scope projects when working with developers

When working with a human developer, the best thing to do is to give them your vision for the whole project (the big picture) and then plan the milestones together.

When you work with an AI dev tool like Replit, I tried that to start with and it really didn’t work. With AI dev tools like Replit, you need to just give them a very defined, narrowly scoped MVP and then add small features on top of that afterwards.

However, the problem with that is that you can end up with a kind of a house of cards situation where one additional feature can break the whole product. I’m still trying to figure out the right way to approach this, but it’s definitely quite different to how you would traditionally plan software.

2 - I was building with AI and building an AI product

This was a mistake in hindsight for my first product. I was building the app using Replit and then the product itself also relied on an OpenAI integration to generate content.

So every time something broke, I wasn’t sure:

• Was it the app logic?

• Was it a bad prompt?

• Was it just the model being weird?

It slowed me down a lot, and made debugging messy.

3 - Test-driven development is the way to go

Bugs that show up later often started way earlier. I thought I was testing properly, but really, you need clearly defined tests for each thing you build. And you need to define those tests with the tool itself so it understands what it’s being tested against. Otherwise, it compounds and becomes very difficult to trace.

4 - I spent too long on the UI

I didn’t like the UI Replit generated, so I designed it in Polymet and uploaded it into Replit. That worked at first but then Replit tried to preserve the file structure from Polymet exactly, and that caused a bunch of other issues. In hindsight, I should have just accepted the basic UI and kept it moving.

5 - The illusion of progress is real

These tools are incredible for getting working prototypes. But to actually get a functional product that people can adopt is tough.

What I’m doing instead

This is my first time really trying to do this. So I think in hindsight, instead of building a product that had the complexities of integrating with OpenAI to generate proposals as the core value prop, I’m going to start first by building a basic SaaS product. A product that solves a single use case and doesn’t require AI to be a core part of the value proposition.

Once I’ve done that, I think I’ll have the confidence to go back and build Proposal GPT properly.

Anyway. I didn’t finish the challenge.

Learned loads.

Will try again.