Provide the user limited control over agent execution. Below are examples of step control, context control and prompt control.
Example - Multi-step writing:
When generating documentation, Sonnet wants very badly to write bullet lists. It works best to use a second pass for consolidating bullets to narrative prose. It would be great if the agent could run multi-pass processes in one command.
Other examples:
pruning (write code, then remove X),
context constraint (focus on this category of files for this search)
search tool selection (use grep to search for X, then vector to expand context)
deep search (search for dependents of X, then search for second-order deps of type Y)
meta-prompting / prompt templating (use prompt X, but customize it for Y usage)
I’ve been doing some learnings on Documenting-As-You-Go – You may be interested in my postsw
Specifically though, I tell the Agent all sort of behavior I want from it. - Check this out (Im gunna do this for the first time, lets ee how it works out:
Thanks @SoMaCoSF , I’ll check your posts.
I found I was able to improve the design ability of Agent by having it build a technical design doc as informal proof. I’ll put more detail on one of your threads.
