← Back to context

Comment by tarboreus

2 months ago

I was thinking about doing something like this. I write a lot of boilerplate about "ok we need to discuss before you implement, ask clarifying questions" so it doesn't go rushing ahead.

yea this boilerplate is super simple and effective. It's essentially mimicing how I use claude code and cursor, but with low level control.

A very cool thing I'm working on in this space is having an llm code with the codesys sdk, then run the code.

So imagine cursor coding a codesys file instead of doing the task directly so that it instead scripts claude code to do a sequence of actions and allows cursor/the user to simply analyze the results.

this also enables parralel claude code sessions which is super cool!

same but the interaction between me and Claude is too dynamic, and my planning framework is more robust than single files. If anything I could speed my workflow up but writing a bash script that calls gemini for my plans - i feed entire codebases or parts of a large codebase (using Prompt Tower) to gemini for the planning - Claude Code isnt as reliable. But even that is iterative with Gemini.

  • I haven’t used Gemini as much but it doesn’t seem as smart as Claude, especially at exploring the codebase.

    I’ve never heard of prompt tower either, so definitely something to look into!