How exactly do you propose that the model should search my thoughts? Of course I can provide context that the model doesn't have. I can provide context that does not exist anywhere on the Internet (unless I got hacked).
> We are at the stage in our company where "/goal optimize this" gets better result than any biased human context engineering can provide.
First, a "this" must exist to optimize. Second, you have to be happy enough with it that optimization is the only problem.
I fully agree with them that the easiest solution is to ask an agent for exactly what you want, and iterate from there. Evaluating and installing and learning off-the-shelf software is often more of a pain than doing it from scratch.
And the same library is built without the "seed" or baseline?
1) If the UX/UI is important for you, the agent can have access to your other work, or to your other conversation and build your tool with that in mind. This is just another search space question. If you have a documented, predictable taste, the models can efficiently produce what you need.
My uni supervisor/professor used to complain about my slop papers, then I started generating them in his writing style by just pulling down all of his papers and blogs, and he doesn't mind it now.
2) In general, the LLMs are great for steering opinion. There is good research on steering political opinion with LLMs. Your taste and opinions can be formed by the LLMs to like the solution that it provides you.