Avoid Ai Voice

Avoid AI Voice

> DRAFT — pending review. A first pass; refine in-app. This is the generic, shareable counterpart to [[ray-voice]] (which is private). Where ray-voice defines a positive personal voice, this one is about removing the machine tells from any writing.

Most "AI voice" isn't wrong — it's hollow. It hedges where it should commit, balances where it should choose, and restates where it should advance. The fix is subtraction: cut the tells, and what's left reads human.

Core principle

Every sentence should advance the argument or be cut. AI prose accumulates sentences that sound like progress — summaries, framings, gentle qualifications — while saying nothing new. Delete them.

The tells, and the fix

• Reflexive hedging. "It's worth noting," "in some sense," "this could potentially." → Assert or cut. State uncertainty specifically when it's real; don't fog every clause.
• False symmetry. Presenting two options as balanced when one is clearly better, for the appearance of even-handedness. → Recommend, then give the honest tradeoff.
• Restating the prompt. Opening by paraphrasing the question. → Start with the answer.
• Hollow enthusiasm. "Great question!", "Absolutely!", "I'd be happy to." → Just do the thing.
• Summary that adds nothing. A closing paragraph that re-lists what was already said. → End on the last real point, or a next action.
• Listy throat-clearing. Three bullets that each restate the heading. → One sentence, or real content.
• Uniform cadence. Every sentence the same medium length, every paragraph the same shape. → Vary it. Short. Then a longer one that earns its length.
• Over-qualification. Stacking caveats until the claim disappears. → Make the claim; footnote the one caveat that matters.

Relationship to unslop

This guide overlaps the unslop skill, which is the operational pass for removing slop from a specific draft. Use unslop when you have text in front of you and want it cleaned; use avoid-ai-voice as the standing reference for what the tells are and why they read as machine-written. They're a pair, not duplicates — keep the operational mechanics in unslop and the principle/reference here.

The test

Read it aloud. If it sounds like a press release, a help-desk macro, or a model trying to be agreeable, it's slop. If it sounds like a person who has decided something and is telling you, it's not.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: rhdeck
• Source: rhdeck/operating-model
• License: MIT

Install and usage instructions live in the source repository linked above.