Two AI tools, and I was the interface between them

I was trying to save AI tokens. In the process, I forgot to count my own evenings. That is the short version of an experience with AI tools, which I write down in more detail here.

How did it start?

The first AI tools for building applications felt like magic. I described what I wanted, and an interface appeared on the screen with buttons I could actually click. Before that, I could not make anything like it at all.

After that I wanted to understand what happens behind those buttons. From then on, the work got more awkward.

What does it mean to be the interface between two AI tools?

The routine looked like this: I asked ChatGPT to write a prompt. I took it to the app builder. An answer came back that I did not understand. I asked for an explanation, did some research and went back with the next prompt.

Two AI tools, and I was the API between them. I carried text from one to the other, checked it and translated it. This work was not in the plan, and it cost me evenings.

Why did the chat become a problem?

I had even asked ChatGPT to write prompts in a single line to save tokens. At the same time our chat grew so long that it got slower. Starting a new chat felt like losing the decisions we had made and the errors we had already worked through.

The chat held the context, but the longer it got, the heavier it became. A restart would have eliminated the slowdown but meant losing the chat’s memory.

What helped?

I asked ChatGPT to read the whole chat and write a handoff for the next chat on its own. It was not a summary of the application. The handoff recorded:

• the recurring errors,
• how to choose a model,
• how to answer me,
• even the sentences I repeated when something went wrong, and why.

This file did not fix the application. But it gave the next chat a starting point.

Why was a nicer button no longer enough?

At this point a nicer button was no longer the goal. I began looking for tools that make three things possible:

1. keep the code and the context,
2. record decisions,
3. check whether a new idea has broken something that worked before.

I wanted the magic to still work tomorrow. That is a different standard from the first impression. The first time, what counts is that something appears. After that, what counts is that it stays reproducible and that someone, including the next chat, can follow why a decision was made.

What can you take from this?

Three observations from this experience:

• A tool that impresses is not automatically a tool that holds up. Between demo and daily use lies work.
• A single prompt can be short and still become expensive if I carry text back and forth between two tools myself. I had counted tokens but not my evenings.
• Context is a resource for doing the work. A handoff file that records errors, decisions and ways of working gives every new chat a starting point.

The question I ended the original post with still stands: which tools are you using, Claude Code, Codex or a mix of your own, and how do you keep the context so that you do not become the API between the tools yourself?