One of the big problems with trying to get just code out of an LLM is that by doing so, you hamper the ability to generate context that assists in the task (aka ReAct).
Also the narrative description of the problem up top is a bit better description than the prompt. Based on my experience doing similar tasks with code, it’s easy to get bad results by setting the task in terms of how to do it rather than describing what you want.
Has anyone managed something similar with an LLM that they can run locally? Given the privacy policies of most of these services, I don’t really want to upload my code to them.
One of the big problems with trying to get just code out of an LLM is that by doing so, you hamper the ability to generate context that assists in the task (aka ReAct).
Also the narrative description of the problem up top is a bit better description than the prompt. Based on my experience doing similar tasks with code, it’s easy to get bad results by setting the task in terms of how to do it rather than describing what you want.
The React code snippets are HTML escaped. This makes it difficult to read some of the code samples.
Has anyone managed something similar with an LLM that they can run locally? Given the privacy policies of most of these services, I don’t really want to upload my code to them.
Definitely check out aider.chat, which automates this process of running ChatGPT and applying diffs.