Keep thinking, keep the vector of your current thought sharp

Coding and working with agents means reading and thinking about a lot of generated material. There's something about talking with an LLM turn after turn that grinds you down and makes you want to turn off your own brain and accept the last output. When you do, the quality starts diminishing almost immediately.

What goes into the context is a signal, and the model amplifies whatever signal it's given. Context is scarce, in tokens, but more importantly in how pointed the information is that you fill it with. There is a clarity of mind signal that happens when you challenge yourself to write with a true voice and say what you actually mean. My experience has been that when the quality of this signal going in is strong and clear, and that quality can be maintained over turns, results are best and worthwhile.

This is also why it's extremely useful to actually know something about what you're talking about. Without knowledge, breadth, or life experience behind what's going in, the output will be predictable and boring.

There is a kind of endurance training here: holding a thought direction, a vector of motion in your mind, through the world of ideas.