| Github | https://github.com/st3fan |
| Blog | https://stefan.arentz.ca |
| Github | https://github.com/st3fan |
| Blog | https://stefan.arentz.ca |
@jasongorman I don't think that is actually correct. I am pretty sure that Claude will use your CLAUDE.md files when you start a new session or when you /clear or /compact - i will test this but I am pretty confident they become part of the always present system prompt.
I do use long running sessions but work in smaller features / changes .. I find it helps a lot to keep a lot of context alive. (Also costs more tokens)
First, I am not vibe coding - I work more transactionally and read all code.
So I have looked at the code that is generated both with and without instructions added to the repo or globally.
It is probably not a big surprise that my guidelines are not being followed when they are not present. This results in code that does not meet my standards.
I am sure there is some overhead like the paper mentions. But my personal experience is that instructions actually do help in a big way
@jasongorman Smaller than "Add a function that ..." ?
I think for me this research is falling apart. It just doesn't match my reality of working with coding agents. I'm getting considerably better results when I add some instructions to the repo. The code and process is much closer to what I want it to be.
(Re smaller steps - it is not a problem really - i'm also having great success with fairly large plans / tasks)