The thinking is yours, the models do the typing

The most useful frame I've found for working with language models is this: the thinking is yours, the models do the typing. The reason that frame matters isn't moral — it's practical. The part of the work where you bring something only you can bring is the slowest part. The part the model does faster than you is the part where you already know the answer. Confusing those two is the source of most of the regret people have about AI tools.

Where the regret comes from

When I see someone unhappy with how an AI tool has reshaped their work, it almost always traces back to the same thing — they handed over the part of the job that was actually theirs, and kept the part that was actually the machine's. They asked the model to figure out what they thought. They asked the model to decide what the writing was for. They asked the model what their argument was. And then they typed the part that came back.

The honest version of the loop runs the other way. You decide what the piece is for. You decide what you're arguing. You decide what stays in and what comes out. The model types — fast — what you've already worked out.

What this means for the tools

Most AI chat tools are optimised for the loop in the wrong direction — they push you toward asking the model to think, because the surface area they expose is "type a question, get an answer". A few things change when you take the other framing seriously:

• You want to see the model choice. Different models think differently. When you're asking for typing, that's fine — pick the fast one. When you're cross-checking your own reasoning, you want a different mind. Tools that hide the model are betting you won't notice.
• You want to branch, not overwrite. A conversation thread is a thought in progress. If the second turn ruins the first, the first should still exist. Branching is a thinking technique more than a UI feature.
• You want your instructions to travel with you. The way you frame what you're doing should not have to be retyped every time you switch from Claude to GPT to Gemini. If it does, the tool is making you do the carrying. That's what oran.chat's shared instruction set exists for.

The thing to keep

I keep this idea on a sticky note above my desk. The thinking is yours, the models do the typing. If the tool I'm using makes that hard, I switch tools. If the conversation I'm in makes that hard, I branch and start fresh. If the draft I'm reviewing was thought by something other than me, I don't ship it.

You can use AI to type the most pleasant cover letter of your life. The question that decides whether it was worth doing isn't "did the model help me write it" — it's "did I write the part only I could write".

If the answer is yes, the model helped.

If the answer is no, the model replaced.