divergelisted

# Diverge Interrupt the default path of jumping to the most probable — and least creative — solution. ## Heritage and scope An original Open Science Skills workflow grounded in **Creative Preference Optimization** (Ismayilzada et al., 2025; background in [`references/creative-preference-optimization.md`](references/creative-preference-optimization.md)). Standard preference alignment (RLHF/DPO) optimizes for the most human-expected output, which is by construction the least surprising one. The paper's most accessible remedy — its own "brainstorm-then-select" baseline — needs no fine-tuning: force divergence before convergence, require that at least one approach is surprising and one is novel, and defer quality and implementation until after selection. Use it wherever more than one non-obvious solution exists — creative, architectural, or analytical work — and not for rote tasks with one correct answer (fix this syntax error). When the task itself is still underspecified (goal, constraints, success criteria unsettled), interview before diverging. Matt Pocock's `grill-me` (see [`RECOMMENDED.md`](../../RECOMMENDED.md)) resolves the decision tree the approaches must answer to. Grilling settles the question, and diverge generates genuinely distinct answers to a settled one. **Model.** No separate model call: this runs in whatever model and reasoning effort the current Codex session is already using, not a fixed pin. `$diverge-codex` spawns a fresh subagent for a clean context