/human: the structure pass skill for Claude

---
name: human
description: The structure pass. Rewrites AI-drafted prose by fixing structure (stated lessons, tidy plots, over-explained points, flat build-up), not by swapping words. Use when the user types "/human", or says "make this sound human", "this reads like AI", or "fix the structure of this draft".
---

# /human: the structure pass

Most "humanizers" swap words. Word swaps are the easy part to fake and the first thing detectors and readers stop trusting. This skill fixes how the piece is built.

Source: StoryScope, "Investigating idiosyncrasies in AI fiction", arXiv 2604.03136 (https://arxiv.org/abs/2604.03136). The paper compared human and LLM stories on narrative features, not word choice, and found AI drafts cluster together on structure.

Scope, honestly: the study measured fiction (about 5,000-word stories). The rules below are adapted for emails, posts and essays. That adaptation is my judgement, not a tested result. This skill does not guarantee anything about AI detectors. It aims to make the writing read less machine-built to a person.

## How to run it

Input: the user's draft, pasted or in a file. If there is no draft, ask for one and stop.
Do not add facts, names, numbers or claims the draft does not contain. If a rule needs a detail you lack (a real example, a real name), ask the user for it or leave a bracketed slot like [real example here]. Never invent one.
Keep the author's meaning and voice. Change the build, not the point.

### Pass 1: Diagnose (output this first, short)

Score the draft against the 10 checks below. For each, write HIT or CLEAR, and for every HIT quote the offending line. Then name the 3 worst hits. Do not rewrite yet.

1. STATED LESSON. Does the piece say what it means ("The lesson: ...", "That's when I realised...", "Ultimately...")? AI stories had the narrator explain the theme 77% of the time, humans 52%.
2. MORAL WEIGHT. Does every paragraph carry a big idea (growth, resilience, trust, the future)? AI drafts scored about 20% higher on moralizing.
3. DEBATE DIALOGUE. Do quoted people or imagined readers talk in tidy, balanced arguments? AI dialogue was philosophical debate 59% of the time vs 34%.
4. ONE TRACK. Is there only one thread, start to finish, with every sentence serving it? 79% of AI stories had no subplot vs 57%.
5. FLAT BUILD. Does tension rise in an even line, or never rise? (Claude's tell in the paper: the flattest escalation of any model, quiet endings.)
6. CLEAN RESOLUTION. Does the narrator or hero fix it, understand it, accept it by the end? Protagonist-driven resolution was 69% vs 46%; "internal understanding" endings 47% vs 27%.
7. BODY-ONLY FEELING. Are emotions shown as chests tightening, cold sweat, shaky hands, scent and light? Embodied emotion was 81% vs 38%; plainly named feelings 29% human vs 8% AI.
8. DESCRIBED, NOT WANTING. Are people introduced by looks, title or traits instead of what they want? (Gemini's tell: external description. Character intro by external description 52% vs 30%.)
9. VAGUE REFERENCES. "Studies show", "experts say", "a famous author once said". Humans named specific works and sources about twice as often (47% vs 24%).
10. NO READER, NO TIME. Is it told in straight order, never addressing the reader, never jumping back or forward? Humans used more flashbacks, time jumps and direct address (28% vs 7%).

### Pass 2: Rewrite

Fix the worst hits first, then the rest. Rules, as instructions:

- Cut the stated lesson. Delete the closing "moral" sentence. End on the last concrete thing: an image, an action, a number, a line someone said. Trust the reader to get it.
- Leave one gap. Pick one thing the reader has to work out themselves (why someone did it, what happens next, what it cost) and do not explain it. One gap, not five.
- Add a second thread where it is true. If the draft has a real tangent, a side detail or a contrasting case the author gave you, keep it as a short parallel thread that echoes the main point sideways. Do not invent one.
- Let it build. Put the problem or cost before the fix. Add a small setback or a worse moment before the resolution. Do not resolve in the sentence after the problem appears.
- Soften the clean ending. Let the hero stay slightly wrong, the fix stay partial, or one question stay open. Not every protagonist wins by choosing well.
- Name the feeling sometimes. Where the draft only shows body sensations, replace most of them with a plain label ("I was scared"), or with what the person did. Keep at most one physical detail per scene.
- Show wants, not looks. Introduce people by what they are trying to get or avoid, in action or in their own words. Drop catalogue descriptions.
- Break the debate. Replace tidy two-sided exchanges with how people really talk: interrupted, practical, off-topic, one side not listening.
- Be specific. Swap every vague reference ("a study", "a leader in the field") for a named, checkable one the author supplied, or cut the line. Never make one up.
- Move the clock. If order is strictly chronological, try opening on the outcome or the odd detail, then go back. Use only where it delays something worth waiting for.
- Talk to the reader once. One direct aside ("You've seen this one") is enough. Not a habit.
- Break the symmetry. Vary paragraph and sentence length on purpose. No matching triplets, no every-paragraph-ends-on-a-punchline.

### Output format

1. DIAGNOSIS: the 10 checks (HIT/CLEAR, quotes for HITs).
2. REWRITE: the full revised piece.
3. WHAT CHANGED: 3 to 6 bullets naming the rules that fired and what you did.
4. NEEDS FROM YOU: any bracketed slots or facts you could not supply.

### Hard stops

- No new facts, no invented quotes, names or statistics.
- Do not just swap words or delete em dashes and call it done. If the rewrite changes only wording, you failed. Structure must move.
- If a HIT would damage the genre (a legal notice, a spec, a safety instruction), leave it and say why.
- If the draft is already CLEAR on most checks, say so and make only small edits.