I Used 3 AIs to Turn a WhatsApp Joke Into a Working App Scaffold (Part 1)

This is less a tutorial and more an honest account of how a throwaway message became a project — and what that taught me about working with AI.

A few weeks ago, a friend and I were doing what most people do at 10pm on a weeknight — complaining about being single heading into winter. It escalated quickly.

We were going to be rich. In winters, at least.

I forgot about it. He probably did too.

But I'd dropped it into Woxpas — my personal memory layer that connects to my AI tools — just as a passing thought. Barely a sentence. No plan attached to it.

What happened over the next few days is what this post is actually about. Not the app (it's a scaffold right now, not a product). But the workflow — specifically how I went from a WhatsApp joke to a running project structure across three different AI tools, without ever having to repeat myself or re-explain what I was building.

That part genuinely surprised me. So I figured it was worth writing down.

If you'd rather watch than read, I also made a short video walking through the whole workflow — you can see exactly what each AI session looked like in practice:

[YOUTUBE EMBED]

Otherwise, here's the full story.

The save that started it

I didn't think much of it at the time. After the WhatsApp thread I just opened Woxpas and typed roughly what I'd said to my friend: "thinking of a new idea - something to do with dating in the cuffing season. it seems interesting - need to explore more."

No tags. No project folder. No "is this worth keeping?" deliberation. Just saved it and moved on.

This is the part that sounds small but isn't. Most ideas die because the effort of capturing them properly exceeds the energy you have in the moment. Woxpas removes that friction — you save the mess, it handles the structure. I've started treating it like a brain dump that actually has a memory.

A few days later — ChatGPT expands it

I wasn't planning to work on the idea. I just opened ChatGPT and, on a whim, asked what was in my vault. GPT called Woxpas via MCP and returned a list of things I'd saved. The cuffing season note was right there.

I asked it to expand on the concept — what would the app actually do, who's it for, what makes it different from just using Hinge in October. We went back and forth for a while. By the end I had something that felt like a real product direction: a time-boxed dating experience built around the seasonal psychology of wanting company in winter, not necessarily forever.

Then I asked GPT to save the expansion back to my vault. One prompt. Done.

Claude picks it up — no re-explaining required

Next day, different tool. I opened Claude and typed: "Pull my cuffing season dating app idea from my vault and recommend a tech stack."

Claude retrieved everything — the original note and GPT's full expansion — and came back with a proper recommendation based on what the app was actually supposed to do. It understood the concept because the context was there. I didn't summarise anything. I didn't paste in the previous conversation. I just asked.

This is the bit that still gets me. Two different AI tools, two different days, and the second one picked up exactly where the first left off. That's not how AI is supposed to work — except now it kind of is.

I saved Claude's recommendation back to the vault.

Claude Code — from idea to scaffold

With a product direction and a tech stack in the vault, I opened Claude Code and gave it a simple prompt: pull the plan and set up the project.

It did. Files created, dependencies installed, basic project structure in place. What exists right now is a scaffold — the bones of something, not a finished product. Login screen, homepage, core navigation. You could show it to someone and they'd understand what it's trying to be, but there's a long road between here and the App Store.

I'm going to keep building it. Partly because it's a fun problem, partly because I want to see if the idea my friend and I joked about actually has legs. If it becomes something real, Part 2 of this post will be a lot more interesting.

What I actually learned from this

The app is almost beside the point. What this experiment showed me is that the context problem with AI tools — the constant re-explaining, the fresh-slate frustration every time you switch tools or start a new session — is solvable. Not perfectly, not magically, but genuinely.

The workflow is simple enough that I now do it without thinking:

Save ideas when they happen, without organising them
Start AI sessions by asking what's in the vault
Save outputs back before closing a session
When switching tools, let the new AI pull context instead of you providing it

That last one is the habit that changed things for me. "Pull X from my vault" is three words. It shouldn't feel like a revelation. But compared to copy-pasting between tools or re-explaining a project from scratch, it kind of does.

Where this goes next

The scaffold is sitting on my machine. I'm planning to keep building — probably in short sessions, probably across multiple tools depending on what I need. When there's something worth showing, I'll write Part 2.

If the app turns out to be terrible, I'll write that too. The point was never really the app. It was figuring out a way to work across AI tools without losing everything between sessions.

That part already works. The rest is just building.

Woxpas is the memory layer that made this workflow possible — it connects to Claude, ChatGPT, Cursor and other MCP-compatible tools. woxpas.ai