Creative Diversification — Development skill for Claude Code

Creative Diversification

A Claude Code skill that breaks LLMs out of modal outputs using three escalating techniques — Verbalized Sampling, Recursive Diversification, and a stacked combination of both. It always asks how weird you want to get before generating, and offers escalation at every round.

When you ask an LLM for "10 ideas," you get 10 versions of the same idea. The model knows thousands of valid alternatives — standard decoding just can't reach them. This skill fixes that, with a dial.

The Three Modes

The skill opens every session by asking:

How weird do you want to get?

  CREATIVE — long-tail but grounded. Usable, refined, still in-vocabulary.
  WEIRD    — domain-primed. Textured, unexpected, specific gravity.
  FERAL    — maximum. Primes + tails + probability dial stacked.

After presenting results, it always offers: *more at this level · go weirder · dial back · go deeper on #N · hybrid · constrain to [domain] · done*. No flags to remember — the loop is the skill.

Mode 1: CREATIVE — Verbalized Sampling

Based on Zhang et al. (2025). Asks the model to verbalize a probability distribution over candidate responses, then thresholds to the tail. The probability threshold is a tunable diversity dial (0.10 → 0.001). Produces long-tail ideas that remain in-vocabulary. Fast, zero-overhead, Pareto-optimal on quality/diversity.

Mode 2: WEIRD — Recursive Diversification

Based on Luo, King et al. (2026). Prepends a random domain lens (*"Through the lens of [DOMAIN]:"*) and injects a random 3-letter stem constraint. Redirects generation toward entirely different knowledge regions. Produces domain-textured results with specific gravity borrowed from crafts, ecology, disciplines.

Mode 3: FERAL — VS + RD Stacked

Uses RD's prime+stem to force domain travel, then applies VS's probability threshold within that primed space. The densest long-tail coverage of any method — more grounded than VS alone, more tunable than RD alone.

The Proble