Himawari Track 2 — Single-Shot Style Caption Agent Docker agent: /input/tasks.json → /output/results.json with four style-specific captions per clip. Approach: describe → single specialist caption (no selector) Sample evenly spaced frames across the clip Claude Sonnet writes a neutral factual description (visible details only) One specialist call per style sees frames + description and returns one caption (personas, tone guides, off-scene exemplars) No best-of-N and no selector—the specialist output is the submission Why: Judges score accuracy × unmistakable style. A selector over candidates tends to pick “safe” voice and blunt sarcasm/humour. Single-shot keeps each style’s sharp write while frames + description keep claims grounded.
Category tags: