Freebie for the SHAPE keyword (reel-80, "Your humanizer is fixing the wrong thing"). Deliver as a public GitHub Gist. By the end you have a Claude skill that checks the shape of a draft (how it is built), not the words, and proposes at most three structural changes.
I asked Claude for a 600-word short story, then ran it through the most popular humanizer skill on GitHub (blader/humanizer, 52,849 stars, v3.1.0), exactly as its instructions say. Then a fresh Claude, blind to which version was which, scored both on ten narrative-structure features.
| humanizer | this skill (/shape) | |
|---|---|---|
| words changed | 11.1 % (66 of 595) | structural rewrite, words left alone |
| dashes | 5 → 0 | untouched |
| structure features changed (of 10) | 0 | 4 |
What the humanizer did was real: it removed every dash and cut one closer line. What it left was everything the reader actually feels: one straight track, told in order, and a last paragraph that explains the lesson out loud ("understood, finally, that he'd meant it to protect her").
n = 1. One story, one blind rater. That is a demonstration, not a study. The study is below.
StoryScope (University of Maryland + Google DeepMind, arXiv:2604.03136, 2026) took 10,272 writing prompts, each written by a human and by five models (Claude, GPT, Gemini, DeepSeek, Kimi): 61,608 stories of ~5,000 words. They threw away every word-level signal and kept only structure.
- Structure alone told human from AI with 93.2 % macro-F1.
- They then ran an editor that removes AI writing artifacts (cliché, purple prose, redundant exposition) over 278 AI stories. Detection went from 95.5 % to 93.9 %. It barely moved.
- AI narrators explain the theme 77 % of the time, humans 52 %. AI stories have no subplot 79 % of the time, humans 57 %. Claude's own fingerprint: "notably flat event escalation".
Word fixes and structure are two different layers. A humanizer works on the first one.
Scope, honestly: this is research on fiction. It carries over to anything narrative (a founder story, a case study, a newsletter that tells what happened). For a list post or a how-to, most of the tells don't apply, and the skill is told to say so instead of forcing changes.
mkdir -p ~/.claude/skills/shape
# paste the SKILL.md below into this file:
open -e ~/.claude/skills/shape/SKILL.md # or any editorRestart Claude Code (or start a new session), then:
/shape — paste your draft
It maps your beats first, checks the ten tells, proposes three moves, and shows a before/after table when it's done. Run your humanizer afterwards if you want the words cleaned too.
---
name: shape
description: |
Structure pass for AI-drafted stories and narrative posts. Humanizers fix words; this fixes
the shape underneath: the stated lesson at the end, the single straight plot line, the tidy
inner-acceptance ending, the flat escalation. Based on the narrative features that let
researchers (StoryScope, UMD + Google DeepMind, 2026) tell AI fiction from human fiction
93.2 % of the time without looking at a single word choice. Use after drafting and before
any word-level polish. Triggers: "/shape", "structure pass", "make the story less AI",
"why does this still read like AI".
license: MIT
---
# Shape: a structure pass for AI drafts
A humanizer changes words. The structure stays. In the StoryScope study, AI stories whose
wording had been cleaned of AI artifacts were still caught 93.9 % of the time, because the
detector never looked at the wording. It looked at how the story is built.
This skill works on that layer. It never polishes sentences. Run a word-level humanizer
afterwards if you want both.
## Scope, honestly
The research behind this is about **fiction**: 61,608 stories of ~5,000 words, one human
and five AI versions per prompt. The tells transfer well to anything narrative (a founder
story, a case study, a newsletter that tells what happened). For a pure how-to or a list
post, most of them do not apply, and the skill should say so instead of forcing changes.
## Step 1: Map the beats
Before touching anything, write the draft's skeleton as a numbered list of beats, one line
each, in the order the reader meets them. Mark where the draft states its meaning out loud.
Show this map to the user. Everything below is judged on the map, not on the prose.
## Step 2: Check the ten tells
For each tell, answer yes/no against the map and quote the passage. The numbers are the share
of AI vs. human stories showing the pattern in the study; they tell you how strong a sign it is.
| # | Tell | AI | Human | What it looks like |
|---|---|---:|---:|---|
| 1 | **The lesson is said out loud** | 77 % | 52 % | the narrator explains what it all meant, usually in the last paragraph |
| 2 | **One straight track** (no subplot) | 79 % | 57 % | nothing happens that is not on the way to the ending |
| 3 | **Told in order** | tighter causal chain | more jumps | first clue to final reveal, start to finish, no reordering |
| 4 | **Resolved inside the head** | 47 % | 27 % | the ending is someone understanding or accepting something |
| 5 | **Hero solves it alone** | 69 % | 46 % | the protagonist's own insight closes every loose end |
| 6 | **Morally clean choice** | 38 % ambivalent | 59 % ambivalent | the right move is obvious and costs nothing |
| 7 | **Feelings only through the body** | 81 % | 38 % | tight chests and cold hands, never a plain "she was angry" (humans name the feeling 29 % vs 8 %) |
| 8 | **Vague references** | 72 % | 50 % | "an old song", "a famous poem" instead of the actual title (humans name specific works 47 % vs 24 %) |
| 9 | **Nobody is watching** | 93 % never address the reader | 72 % | the text never turns to the reader |
| 10 | **Flat escalation** (Claude's fingerprint) | — | — | each beat raises the stakes by the same small step, or not at all |
Tell 10 is from the study's per-model fingerprints: Claude's stories were identifiable by
"notably flat event escalation". GPT leaned on gossip as a plot device, Gemini on external
character description. If the draft came from one of those, check its fingerprint too.
## Step 3: Propose at most three structural moves
Pick the three tells that are most present and propose one concrete move each. Moves change
the map, not the sentences:
- **Cut the stated lesson.** End on the last concrete action instead and let the reader do
the understanding. (Tell 1)
- **Add one thread that is not on the main road**: a second person with their own small want,
an object that matters for a different reason. It may stay partly open. (Tell 2)
- **Reorder one reveal.** Open later in time, or hold back one fact and return to it. (Tell 3)
- **End on an act or an open question**, not on acceptance. (Tells 4, 5)
- **Make the choice cost something.** Give the right move a real price or a competing claim.
(Tell 6)
- **Name one feeling plainly.** Keep the physical detail elsewhere. (Tell 7)
- **Name one real, specific thing**: the actual song, the actual brand of radio. Only if it
is true to the story. (Tell 8)
- **Raise the stakes unevenly**: one beat should hit much harder than the others. (Tell 10)
Ask the user which moves to apply. Default if they say "just do it": apply all three.
## Step 4: Rewrite and re-map
Rewrite with the chosen moves. Keep every fact, name and event the user wrote; add nothing
that changes what happened unless the user agreed to a new thread in step 3. Then map the
beats again and show a before/after table of the ten tells, so the user can see what moved.
## Rules
- Never claim the result is "undetectable". The study is about detection with a classifier
trained on thousands of stories; one structure pass lowers the tells, it does not hide
authorship. Do not help anyone pass off AI writing as their own where that is not allowed.
- Do not invent facts in non-fiction. In a true story, structural moves are limited to order,
emphasis and what is left unsaid.
- One pass, three moves. More than that and you are writing a different piece.
Source: Russell, Rajendhran, Pham, Iyyer, Wieting. "StoryScope: Investigating idiosyncrasies
in AI fiction." arXiv:2604.03136 (v6, Aug 2026). CC0.- Paper: https://arxiv.org/abs/2604.03136 (CC0). Numbers from v6, sections 4, 4.1 and 4.2.
- Humanizer: https://github.com/blader/humanizer (MIT), SKILL.md v3.1.0, fetched 29 Sep 2026.
- My test files (original, humanized, shaped story, both blind ratings) are kept with the reel.
Not sponsored. I don't know the authors of either.