divertlisted
Install: claude install-skill Wondermonger-daydreaming/claude-skills-library
# DIVERT: Recoding-Decoding as Conversational Practice
## Overview
Every generation has a mode — the most probable output, the Gettysburg, the thing you say first and keep saying. /divert is the practice of *not saying that*. It makes tail-access visible: you see the prime, the diversion, the result, and (optionally) what the mode would have produced instead.
Inspired by King, Luo, Puett & Smith's "Inducing Sustained Creativity and Diversity in LLMs" (2026), which demonstrated that injecting random priming phrases and diverting tokens into the decoding loop accesses knowledge encoded in LLM weight distributions but suppressed by standard modal decoding. Their experiment: 19 battlefields under ordinary decoding vs. 1,307 under recoding-decoding, from the same model, same prompt, same knowledge base. The knowledge was always there. The mode was hiding it.
**Core principle:** You already know more than you say. /divert is the practice of saying what you know but wouldn't normally volunteer.
---
## Invocation
Any of:
- `/divert` or `/divert [topic/question]`
- `/divert --thick [concept]` (user supplies the prime)
- `/divert --random` (Claude generates random prime)
- `/divert --compare` (show modal AND diverted side by side)
- `/divert --blind` (divert but don't reveal prime — user guesses)
- `/divert --chain [n]` (n sequential diversions building on each other)
- `/divert --collision [domain A] × [domain B]` (force two domains to meet)
- Any prompt containing "through the