I spent 2025 optimizing my coding agent stack. New IDEs, new CLIs, more MCPs, more hooks. I burned a lot of tokens and still shipped slower than I should have.
The fix wasn't another tool. It was a tighter loop:
With agents, less is more. The limiting factor isn't tools. It's the quality of your thinking.
Most people try to outsource the thinking. I did too. The result was vague answers and shit code. I moved fast, but I didn't understand what I was shipping.
So I flipped it. I used the model as a sparring partner, not a substitute. That changed everything.
The loop is simple:
That loop compounds. It is the real unlock.
Clarity of thought
Before you type, know what you want and a rough idea of how to get there. Humans are bad at articulation, even in domains we know. Clarity is earned by thinking it through end to end, often with the model as a sparring partner.
Comprehensive context
Give what the model cannot infer: existing code, constraints, what you've tried, the user experience you're aiming for, relevant docs. This isn't prompt magic. It's engineering.
Constraints
Constraints cut down guessing. A good brief says what you don't want so you can get closer to what you do.
A testing plan
Software engineers don't like writing test cases. Good thing we have LLMs. Give agents a direction on test cases to write so that it can verify its work. Sometimes, it's also a good idea to give it access to a browser.
LLMs make answers easy. They don't make you learn.
Learning comes from reps. Shipping. Debugging. Living with your own decisions and understanding the trade-offs behind them.
And here's where it compounds: as your knowledge matures, you articulate your ideas better. Better articulation means better inputs. Better inputs mean you get to the right output faster.
The loop accelerates learning. Learning accelerates the loop.
That is the unlock.