Problem
AI is non-deterministic - like rolling dice. You want a three, but probably won't get it on the first try.
What do you do? Roll five dice.
Pattern
Run multiple implementations in parallel from the same checkpoint:
- Create checkpoint (save plan + git commit)
- Fork into parallel working directories (use git worktrees or similar)
- Launch multiple AI implementations simultaneously
- Review all results
- Pick the best or combine elements from multiple attempts
This is trading tokens (relatively cheap) for your time (expensive).
Two complementary modes:
- Failure Mitigation: Complex feature, uncertain approach. Run 3-5 parallel attempts. Some fail, some succeed. You move forward immediately instead of debugging sequential failures.
- Exploration of the Solution Space: When quality matters more than speed. Generate multiple working versions, compare approaches, combine the best ideas. Works especially well for creative work: UIs, game mechanics, designs.
Examples
Game development (Ricochet Robot): Ran three parallel implementations. First version: no walls, robot didn't move - total failure. Second: mediocre. Third: movement logic worked great, loved the button styling. Combined the working movement with the better buttons.
UI design: Run several parallel implementations. One has great layout, another has clever responsive breakpoints, third has interesting color scheme. Borrow the best from each, combine into richer final design.
Designer collaboration: Designer creates mockup in Figma separately. Run parallel AI implementations. Combine designer's vision with AI's working implementations.
