Sound Human

Sound Human edits or audits AI-assisted writing without flattening the writer's voice. It catches familiar language tics, unsupported authority, repeated machine cadence, and generated-looking product pages.

Relationship to No AI Slop

Version 1.1 adopts the concise edit/audit workflow and self-evaluation pattern established by Peter Yang's excellent No AI Slop, used under its MIT license. Sound Human began as a separate editorial checklist; the overlap became obvious after publication, so this repository now credits the stronger upstream execution model directly rather than pretending parallel invention is a differentiator.

Use Peter's project if you want the established, broadly installed skill and packaged ChatGPT/Codex plugin. Sound Human is a focused extension for:

• evidence, attribution, uncertainty, and manufactured consensus;
• the portability test for generic filler;
• patterns that emerge only through repetition across a full piece;
• product-page structure and visual-design slop;
• operational claims, responsibility boundaries, and invented product artifacts.

What it does

Edit returns a minimally revised draft plus a short note naming the material changes.

Audit quotes each problem, names the pattern, and suggests a brief fix without rewriting or guessing whether AI wrote the text.

Both modes preserve supported facts, citations, uncertainty, personality, humor, and useful rough edges. The skill refuses to invent evidence or make the writer more certain than the original draft.

Install

Claude Code

git clone https://github.com/tymrtn/sound-human ~/.claude/skills/sound-human

Or clone it into .claude/skills/sound-human inside one project.

Hermes-style skill trees

Copy the repository into your skills tree, such as skills/writing/sound-human/, then reload skills.

Codex and AGENTS.md agents

Place the repository in the project and add this instruction to AGENTS.md:

When editing or auditing prose, follow skills/sound-human/SKILL.md and run its eval.md checklist.

Any model

Paste SKILL.md above the draft and ask the model to edit or audit it. For edits, require the model to run eval.md before returning the result.

Files

• SKILL.md: executable editing and audit instructions
• eval.md: self-check before returning an edit
• examples/: before-and-after editorial, product, attribution, travel, and layout examples

Limits

• The skill is calibrated for English editorial, product, and marketing prose.
• A clean pattern audit cannot supply substance the writer never provided.
• The rules identify sourcing problems but do not verify sources.
• Editorial heuristics yield to deliberate voice. Integrity rules do not.

License and attribution

MIT. See LICENSE.

This project incorporates and adapts ideas and workflow structure from No AI Slop, Copyright © 2026 Peter Yang, under the MIT License. Subsequent Sound Human additions are Copyright © 2026 Tyler Martin.