Compress, Don't Expand

Treat the LLM like a code factory and it will act like one. Describe a feature, get an implementation, ship it, repeat. More code, more surface area, more to maintain. Every loop feels productive because something lands at the end of it.

You work in expansion mode. You expand the idea into code, and then you ship the code.

A better mode is compression. You use the LLM to interrogate the idea. Argue with it. Poke at it from angles you would not have found on your own. Write three prototypes nobody ever sees and throw them all away. Squeeze the space of possibilities until the one idea worth building falls out.

That is finding the truth. It is the real work.

My rough split is 80/20 as usual. Eighty percent with the LLM is exploratory like chatting, introspecting, sketching, debating, writing code that will never run, discarding it, starting again. Twenty percent is building whatever survived.

Most will run the opposite ratio. They spend 80% building and 20% thinking, because building is visible and thinking is not. The LLM makes the visible part cheap, so they ship even more code that should never have existed. I am guilty of this too. It is a natural trap. The tools are so good at expansion that it is easy to get lost in it.

The tools are ready. The question is what you ask them.

Compress first. Build only what survives.