Go to AI Studio, select Gemini Pro, give it your code, or describe a problem you want to solve, and then tell it that you want to --> discuss <-- the code. That you don't want it to generate code for you, but that you want to discuss it, that you want it to comment on your code or on how to solve the problem.
This is the best way to get Gemini to be a really good assistant, unless you want to add System Instructions which precisely describe how it should behave.
Because if you just say it should solve some problem for you, it eagerly will generate a lot of code for you, or add a lot of code to the clean code which you provided.
Yeah my current approach is to generate a plan with Gemini Pro, with plenty of back and forths, and then have it write the plan to a markdown file. Afterwards I get either it or another model to follow the plan step by step. Without doing this the results are questionable and often require going back and fixing a lot.
Even limiting its scope can be risky, if it's a complex problem in a niche that's not well-represented in the training data. Been learning Gleam lately, and yesterday when wrapping my head around recursion was starting to give me a headache, I tried to ask Gemini Pro to write a function. It needed to recurse down a tree, perform a function on each child that requires information accumulated from each of its ancestors, and return all nodes in the tree as a flat list.
It returned over 600 lines of code across 3 code blocks, almost all of them commented out for some reason, each with an accompanying essay, and each stuffed with hallucinated and unnecessary helper functions. Apparently Gemini Pro struggles to wrap its weights around recursion more than I do. I just wrote it myself and only needed 26 lines. It's not using tail calls, but hey, my target platform still doesn't support tail call optimization in 2025 anyway.