What problem does it solve? Algorithm-first research quests often drift into unfocused tinkering, premature branch creation, or paper-writing work before the optimization search is properly managed. This Skill imposes a disciplined control layer that turns loose ideas into ranked candidate briefs, promotes only the strongest directions into durable lines, and routes each pass toward explore, exploit, fusion, debug, or stop based on evidence. ## Core Features & Use Cases - Candidate Brief Management: Shapes fuzzy directions into structured briefs covering bottleneck, mechanism, expected gain, and risks, then ranks them on one explicit comparison surface before promotion. - Frontier-Driven Routing: Recovers the optimization frontier via artifact and memory calls, then selects exactly one dominant next action per pass across six internal submodes (brief, rank, seed, loop, fusion, debug). - Durable Record Keeping: Maintains OPTIMIZE_CHECKLIST.md and CANDIDATE_BOARD.md as quest-visible control files, and records every candidate, result, and route decision through artifact and memory APIs. - Use Case: A researcher has three plausible ways to improve a model's training pipeline. The Skill generates differentiated briefs, ranks them, promotes the top one into a durable line with its own worktree, seeds bounded smoke candidates, and records measured results before deciding the next route. ## Quick Start Ask the agent to run an optimize pass on the current quest by refreshing the optimization frontier, ranking existing candidate briefs, and recommending the single strongest line to promote next.
