Request - Your help in improving the default system prompts

[olimorris]

Hello All 👋🏼

In my quest to make this plugin yield the best outputs from OpenAI, Anthropic AND Ollama, I'm seeking your feedback in how we can tweak the default system_prompts (and system prompts in the default_prompts table) in the config.lua file.

Ultimately, I'm trying to get to the level of response that you may observe in GitHub's Copilot Chat in VS Code. Helpful, but overly terse and low in token consumption.

As an aside; the changes I've made to the plugin in recent weeks should allow for Gemini integration. It's low on my todo list but conscious they're silently making really interesting improvements to their models.

[JPyke3]

Hmmmm... Do you think it's worth trying to find the best output across the board? Just from what I have used at least it seems almost always beneficial to tweak system prompts based on the specific AI model. I guess we need to start somewhere but maybe it's a consideration to at least split OpenAI and Anthropic's models at a minimum. Unfortunately due the nature of Ollama it's obviously hard as a user can run whatever they want, so that's maybe less important.

Thought I would throw my hat in the ring anyways, I will spend this WE refining my Ollama prompts and post them up so people can take a look

[olimorris]

Very good suggestion. Perhaps even if it's not at a model level we could easily make it an adapter one for Anthropic and OpenAI. I'll play around in the coming days.

Your Ollama prompts would be very helpful.

[JPyke3]

Yeah I think at an adapter level seems pretty reasonable. Now that I think about it, I am running some gpt-turbo-3.5 prompts with gpt-4o and they've never had any issues. So I think that's a good way to run it