Novelty Hunt banner
rwshiraishi rwshiraishi

Novelty Hunt

Development community

Description

Claude Code skill: search a solution space for genuinely original candidates and score them without killing them. Evidence-based - every mechanism traces to a measured result from the creativity-assessment literature.

Installation

This entry records only its repository, not the path inside it, so there is no exact command to give. Open the source below and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

README

Novelty Hunt

**Find genuinely original candidates while protecting them from the quality filter that normally kills them.** One measured finding at a time: every rule traces to a measured constant, and open questions stay listed as UNANSWERED rather than filled with invented numbers.

[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE) [![Lessons on the books](https://img.shields.io/badge/lessons-11%20%289%20promoted%29-8a5cf6.svg)](references/lessons.md) [![Runtime](https://img.shields.io/badge/runtime-markdown%20%2B%20bash-lightgrey.svg)](#requirements) [![Made by Neon Peach, LLC](https://img.shields.io/badge/made%20by-Neon%20Peach%2C%20LLC-ff8c69.svg)](https://neonpeach.co)

Most brainstorm-then-pick workflows fail twice. The generation step produces paraphrases of the obvious answer. The scoring step buries whatever originality survived. Every mechanism in this skill is built against both failures.

Six phases of novelty-hunt: Phase 0 frames the conventional core and freezes the gate, Phase 1 maps the axes, Phase 2 runs disjoint search moves to archive candidates, Phase 3 gates out junk, Phase 3.5 runs execution tests, Phase 4 scores on separate dimensions, Phase 5 delivers three picks.

**Contents:** [Why](#why-this-exists) · [How it works](#how-it-works) · [Field record](#field-record) · [Quick start](#quick-start) · [Install](#install) · [Using it](#using-it) · [Layout](#repository-layout) · [Self-maintenance](#the-skill-maintains-itself) · [Design notes](#design-notes) · [Contributing](#contributing) · [Maintainer](#maintainer)

Why this exists

Aligned models are mode-collapsed. Preference tuning sharpens output toward the familiar. Measured on a Tulu-70B, the base model retained 45.4% semantic diversity, 20.8% after SFT, 10.8% after DPO. Frontier models produce fewer