I tried out dynamic workflows and didn’t get much out of them at first. I spent today trying to s…
I tried out dynamic workflows and didn’t get much out of them at first. I spent today trying to see what I was doing wrong and learned a bunch.
Mainly, I found it so useful for adversarial agent tasks. I just used it to review a 7 PR stack. Here’s how I approached it:
Claude really likes his own work, and that makes it tricky to get quality reviews, even with other standalone agents.
However, splitting the review into many focused segments with ultracode is killing it for me.
Here’s my process:
1 - Define the task, often reviewing a plan or code diff, and the main things to care about. Think of this as the PR description.
2 - Define the swimlanes for the reviewers. For me this comes in a few flavors, but it’s often correctness, code duplication, safety, maintainability, etc. This is where you should put the most brain power into developing your own. (I like to add a philosophy section in my skill to get the agent in the right headspace.)
3 - Turn on ultracode (I hope you already did this).
4 - Tell the model to review what it has first (the diff, the plan, the doc, etc.), then create a workflow to cut up the reviews as thinly as possible to achieve what you want, and to have them be incredibly adversarial. I often like to say “have them be super mean.”
5 - Then (my favorite) include that each agent must come back with a way to verify their findings.
6 - Finally, end with the main Claude verifying all the findings once the list comes back, then prioritizing and presenting them to you.
My advice would be to copy this tweet, give it to Claude, and add your own flavor and goals for what you’re actually reviewing. So far I have seen this catch bugs, produce cleaner code, and stop me from having 28 date formatters across my codebase.
Quoted @trq212: http://x.com/i/article/2061850535708483585
Key Takeaways
- Imported from X — add takeaways on review.