For me, I've set up two additional agents besides Plan and Build:

• Learn: rather than try to implement everything, responds with code snippets and guidelines to implement the feature I want. This speeds up my coding without taking away my agency over the code or building experience. It has become my favorite go-to mode.
• Reviewer: what it says, reviews the code, categorizes its conclusions into: Critical, High, Medium, Low, Polish. Also doesn't modify the code, just provides brief guidelines to resolve the issue.

Pipeline

Generally I start by describing the feature to the Learn agent, it explores the code and creates instructions/guidelines for me to implement it that account for the existing code base expectations and behaviors. I just dive in and implement it.

Then I use the Plan and Build modes to create test suites for the feature as well as implement any trivial code needed.

Once all is clear and ready, I ping the Reviewer to do its thing and usually there are a few false positives, but it does catch real issues. I then implement any necessary fixes and iterate until code is ready.

Model assignments

• Learn: should assign a pro level model with a good linguistic personality. so far I like using: MiMo-V2.5-Pro, Terra-high, DeepSeek V4 Pro
• Plan: should be a frontier-level model. My gotos are: Sol-high and GLM-5.2
• Build: any good coding model. My current gotos are: DeepSeek V4 Flash 0731, Terra-high, Luna-high
• Reviewer: some frontier-level models (like Sol) aren't good for this because they flag way too many "what-if" issues. I prefer: GLM-5.2, Terra-xhigh, DeepSeek V4 Pro