← ClaudeAtlas

ck-arenalisted

Generate independent candidates for the same task, compare them against explicit criteria, and synthesize a verified result. Use when the user asks to arena a task, explore competing designs, or compare multiple attempts before choosing an artifact.
Wang-Cankun/cankun-skills · ★ 2 · AI & Automation · score 78
Install: claude install-skill Wang-Cankun/cankun-skills
# CK Arena Fan out independent attempts at the same task. Read each candidate, choose a base, incorporate the strongest compatible ideas, and verify the resulting artifact. The artifact may be a design, implementation, document, or analysis workflow. Track six phases: Frame, Fan out, Cross-judge, Pick, Graft, Verify. Use the host's available agent and execution capabilities; the workflow does not require a particular platform or model provider. ## A. Frame The shared task is the candidates' contract. 1. State the artifact, constraints, input evidence, output location, and the observable result that will establish success. Accept caller-supplied grounding, candidate instructions, and output templates, such as a design package requested by `ck-architect`. 2. Derive a short rubric with concrete criteria for this task. The judge receives the rubric; candidates receive the same realistic task and necessary constraints. Keep candidate outputs and model identities out of one another's context. 3. Use the number of candidates requested, or start with two distinct attempts and expand only when the decision warrants it. Honor the user's model choices and the host's configured defaults. Different available model families can add perspective; multiple independent runs of the same model are also valid. 4. Give each candidate a separate output directory. Use isolated checkouts when candidates must modify a repository, or scratch directories for self-contained artifacts. Keep input dat