Despite all the hoopla of “project knowledge” and supposed codebase-wide context, I still find reasoning models do their best when directly provided with files relevant to a problem and nothing more.
I plan to add a tree feature and restore some other features I had in prior versions.
There are probably other tools that don’t require completion API requested but assist in AI enhanced dev workflows.
I actually hate using the AIs on third party platforms/wrappers because you have no idea what garbage they are putting in the system prompt and they have a strong incentive to minimize costs on LLM tokens. The difference with the quality I get just dumping all relevant context into the biggest SOTA model with reasoning tokens maxed out is massive.
Its to the point where if Cursor gets stuck on something I will just tell it to give me a recap of the details of the issue and suggest relevant files and then paste that into the Anthropic Console with thinking tokens set to max.
Comments
If you decide to add a utilities section, please consider FileKitty.
https://github.com/banagale/FileKitty
Despite all the hoopla of “project knowledge” and supposed codebase-wide context, I still find reasoning models do their best when directly provided with files relevant to a problem and nothing more.
I plan to add a tree feature and restore some other features I had in prior versions.
There are probably other tools that don’t require completion API requested but assist in AI enhanced dev workflows.
I actually hate using the AIs on third party platforms/wrappers because you have no idea what garbage they are putting in the system prompt and they have a strong incentive to minimize costs on LLM tokens. The difference with the quality I get just dumping all relevant context into the biggest SOTA model with reasoning tokens maxed out is massive.
Its to the point where if Cursor gets stuck on something I will just tell it to give me a recap of the details of the issue and suggest relevant files and then paste that into the Anthropic Console with thinking tokens set to max.