novelty-huntlisted
Install: claude install-skill rwshiraishi/novelty-hunt
# Novelty Hunt
Explore original candidates, preserve meaningful alternatives, and deliver useful picks. The workflow draws on creativity research and recorded runs; study results inform the method but do not establish its effectiveness across every domain. Constants and citations live in `references/evidence.md`.
## Calibration (read first, always)
**Skip floor.** Do not run this skill when: the ask is trivial (variable names, one-liner fixes), a conventional answer is obviously sufficient, or the user wants speed over originality. A hunt costs roughly 10-25k tokens lightweight, 40-90k full. If the fix is smaller than the hunt, answer directly and say why the hunt was skipped.
**Mode selection.** Lightweight (default): single agent, 1-2 search rounds, persona-panel self-scoring, archive of 6-12 candidates. Full mode only on an explicit trigger: user asks for an exhaustive hunt, a prior conventional fix failed, high-stakes one-shot decision (naming, positioning, architecture bet), or user says "full". A mode named by the user always wins over the trigger list; if the triggers disagree with the named mode, run the named mode and say so in one line. Full mode spawns generator and judge subagents and uses cross-model scoring where configured; read `references/generation-moves.md` and `references/scoring-protocol.md` before starting full mode. Lightweight mode needs only this file.
**Lightweight honesty caveat**: in a single-agent run, persona passes share one context — the s