de-ailisted

# De-AI *Read the prose like a reader, not a writer. Where the rhythm announces "AI wrote this", cut.* *Memory Model role: Reads the destination and recent trail entries for established voice norms; applied as a finishing pass during Improve runs that touch prose. Records realisations about new AI-tells into the trail so they accumulate.* > **Governing principle:** [Observable Autonomy](../PRINCIPLES.md#principle-2-observable-autonomy) — the trail must be honest, and so must the prose. Mannered AI writing erodes the reader's trust that anything beneath it is real. A reader who notices the style stops engaging with the substance. This skill exists because LLMs converge on a recognisable prose register: parallel-structure-itis, em-dash bridges, restatement bloat, listicle bait, mannered meta-framing. The patterns are not wrong individually. They become a tell when they cluster. Once a reader recognises the cluster, they read the text as machine output regardless of how good the underlying content is. This skill is the finishing pass that strips the cluster. --- ## What this skill is not - **Not a search-and-replace dictionary.** Em-dashes are fine. Three-item lists are fine. The patterns below are diagnostics, not banned tokens. A pattern is a tell when it is the default, not when it appears once. - **Not a substitute for Improve.** Improve fixes the substance — what the text claims and how it is structured. De-AI runs after, on the surface. Applying De-AI to text that