I often use Cursor and Windsurf for Vibe coding. When I ask them to fix slightly complex bugs, they sometimes struggle — even after prompts like “think hard” or “fix really”
Here’s a trick that works surprisingly well
- Ask the LLM to only add log lines (no fixes yet) to help debug the issue.
- Re-run the code and share the log output with the LLM. Ask it to investigate and fix the problem using those logs.
- Once it’s fixed, ask it to remove all log lines containing your special tag.
This approach works consistently — the model can actually reason through the issue better when it sees real log context, just like a human would.