Model response / prompt

### Description:

Once we have a kind of more polished version of #20 ready, we have to do some nice testing with the current system prompt and how the model should act. Also there might be some UX improvments as well based on the mode we choose. Currently we are using `deepseek/deepseek-v4-pro` with resoning `medium` but we can tweak / use some other model.

Start by reviewing the polished version of #20 and the current system prompt and model configuration. Compare prompt and model options, including the current DeepSeek model and medium reasoning setting, then evaluate the resulting behavior and UX. Done means a tested choice of prompt/model behavior and any resulting UX improvements are documented.

Written by the indexing model from the issue text.