Do this. Take all 3 prompts in this directory bolt.diy/app/lib/common/prompts and copy and paste each in Qwen3-Max https://chat.qwen.ai/ and ask it to improve it for Qwen3-Coder “flawless uninterrrupted vibe coding on bolt.diy”.
If bolt.diy delivers an error or does not deliver a preview at the first run paste the results in to Qwen3-Max. Do this until you get a 90% success rate. Do not over do this otherwise prompt will get to big depending on you ctx size and max token you run in problems.
You will get a prompt which delivers much more consistent results in one go.
Also you can try this specialized model of Qwen-Coder:
Qwen3-Coder-30B-A3B-Instruct-480B-Distill-V2-Fp32
Take the highest quantization you can run! IQ4, Q4_KS or bigger.