Humanese — AI skill for Claude Code

humanese

**A Claude Code skill that translates dense AI output back into human language.**

If your model's answers started reading like this —

`Done: auth fix deployed. Next: regression suite. Risk: token refresh edge case. No broad refactor.`

— you've met **neuralese**: the telegraphic dialect post-trained models drift into. Every token carries maximum weight, nothing accommodates the reader. It's not a bug and not laziness; it's the optimum of the reward function. Being right and being dense scores points. Being *readable* doesn't.

`humanese` is the leash applied after the fact. Paste any AI-written text (a report from another session, an overnight agent loop, a subagent handoff, a tweet) and it gets decompressed back toward the human anchor — at whatever depth the receiver needs.

The dial

Level	Name	What you get
N1	Glossary	Original intact, acronyms expanded, short glossary at the end
N2	Five Words	Fragments become full sentences; labels and arrow chains become prose
N3	Decompression	Default. Every concept explained on first appearance; facts, decisions, risks and open questions kept separate; hedges preserved verbatim
N4	Milk	Didactic — analogies, examples, step by step. For someone outside the project
N5	Embodiment	Written for ONE named person, in their language, with their examples

Core principle: **the reward is the receiver's clarity, never the size of the output.** Decompression is not summarizing — the output is usually *longer* than the input, and every fact, number, condition and uncertainty survives.

The best side effect: **dense hallucination breaks on expansion.** A sentence that won't open into concrete meaning gets flagged `[OPAQUE]` instead of being paraphrased in the dark. Every decompression is a free audit.

Install

Copy the skill into your Claude Code skills directory:

git clone https://github.com/UDIIA/humanese.git
mkdir -p ~/.cla