mike-diff /best-of-n — Parallel Model Selection
Guidance for using Cursor's parallel attempt feature to improve output quality.
When to use
- Ambiguous requirements where multiple approaches are viable
- High-stakes code where correctness matters (security, data integrity)
- Creative tasks where solution diversity improves outcomes
How it works
/best-of-nspawns N independent attempts at the same prompt- Each attempt runs in its own isolated worktree
- Results are evaluated and the strongest is selected
- Other attempts are discarded
Best practices
- Use N=2–3 for most cases; higher N for critical paths
- Write clear evaluation criteria in the prompt
- Review the selected output before merging — selection is heuristic, not perfect