divergence-loop
Divergence Loop
A brainstorming skill that forces genuine creative divergence by exploiting context isolation between generation rounds, then bridges wild ideas back to reality through constraint-based convergence.
Why Context Reset Matters
LLMs have a strong convergence bias -- given enough context, they drift toward safe, structured, "reasonable" outputs. Prompt-level instructions ("don't converge", "be wild") fail because the model's training rewards coherence over chaos. The only reliable way to maintain divergence across multiple rounds is context isolation: each generation round runs in a fresh execution context with zero memory of prior rounds. Files are the sole communication channel.
How to achieve context isolation depends on the platform. Use whatever mechanism provides a clean context with no carry-over from prior rounds (e.g., spawning a new agent, opening a new chat, or calling an API with an independent message history). The specific tool does not matter — what matters is that each round's generator has NO access to ideas from other rounds.