Stake · Hook · Arc · Prune · Evaluate
A writing method for people and teams working with AI. The engine is open source; your voice stays yours.
Earlier this year, five researchers collected 61,608 stories. One set came from people, the rest from five language models, and they stripped every stylistic clue: word choice, punctuation, rhythm, all of it. Then a classifier had to tell human from machine from narrative structure alone. It succeeded 93 percent of the time. Even with all the word clues removed, the structure gave the machine away.
Your readers run the same experiment every morning, without a lab. They open an article, feel it go flat by paragraph three, and leave. Ask what was wrong and they say it sounded like AI, even with every em-dash hunted down and every "delve" replaced. What they felt was the structure: every point delivered with the same weight, an argument that only agrees with itself, an ending that resolves everything and risks nothing.
The standard response is a word cleanup. A ban list for the obvious AI phrases, a tone instruction in the prompt, an editing pass over the wording. The research above explains why the results keep disappointing: the classifier barely needed words. The even pacing, the argument without friction, the tidy close, all of it survives a vocabulary sweep. The effort goes where the problem was solved, and the giveaway stays where nobody looks.
SHAPE starts from the opposite conclusion: govern the decisions and let the machine write the sentences.
Each step has to be finished before the next may start. Full definitions with gates: SKILL.md.
Stake. Before anything gets drafted, one sentence answers two questions: who is this for, and what changes for that person. A real person you can picture: a CFO defending a budget, or your own board. We call it the stake because it names what that reader stands to win or lose. If the sentence comes out empty, the piece gets dropped an hour in, before it costs a week. And when a machine drafts, this one sentence steers it better than any prompt trick.
Hook. The first line drops the reader into the story at the point where it concerns them. On a feed, the claim comes first, and the warm-up paragraph most drafts open with gets deleted, because the reader decides whether to stay before the warm-up ends.
Arc. Every piece gets told with rising stakes, even a short post. The test needs no taste: if two sections can trade places and nothing is lost, the piece is flat, and flat is exactly what that classifier measures in machine writing.
Prune. When the draft is done, it gets cut back. The wording rules run as automated checks, because a computer never forgets a rule and people always do. Then the structure gets tested. Then ten percent comes off a text that already read well.
Evaluate. The step the other four exist for. After each piece is published, the published version gets compared with the draft it started from. The differences get named as patterns, and each pattern becomes a dated rule with the failing example attached.
One July morning, an editor produced a 174-word announcement post. The author cut it to 50, keeping only the final paragraph, where the argument had been sitting all along. By evening the system carried a new rule: in short form, open on the claim; proof stays in the article. Every draft since starts from that lesson.
Prompt libraries age the day they get written. A rulebook that remembers gets better with every piece, because it learns from your own texts, your own readers, and your own mistakes. That is also why a copied rulebook is worth so little: the lessons in it belong to someone else.
Some weeks the method tells you the idea you loved has nothing in it for your reader, and the piece never gets written. Some days it deletes the paragraph you were proudest of. That is the price, and it is real.
Discipline decays exactly where you would predict, too. Our own editor, with the em-dash ban written down in four files, reintroduced em-dashes during a routine revision, because in that pass the rules sat in nobody's head. We keep that failure in our worked examples on purpose, because it proves the point better than any success: a method that relies on remembering is a mood, but a check that runs on its own is a control.
One thing stays manual, and should. The rules only accumulate if a person decides which lessons deserve to become rules.
| File | What it is |
|---|---|
SKILL.md |
The engine. The five steps as an agent-executable skill (Claude, Codex and compatible agent runtimes), versioned, with a changelog. Carries no house voice and no audience of its own. |
config.template/ |
The three files you fill in once: your readers, your voice reference, your rulebook. Copy to config/; how long it takes is up to you. |
GUIDE.md |
The setup walkthrough: how to describe your voice from your own best work, mine your drafts for writing patterns, and wire the whole thing into your agent. Model proposes, you confirm. |
EXAMPLES.md |
Real before-and-after pairs from our own output, including the failures. The part a method description cannot fake. |
How the split works: the method anchors your voice in your own published work, never in an imitation of a named author. The engine never contains your voice, and your config/ never leaves your machine. You can pull engine updates forever without touching your voice, and contribute engine improvements without leaking it. Where the engine and your rulebook disagree, your rulebook wins.
Using it with an AI agent: point your agent at SKILL.md with your config/ next to it. The skill instructs the agent to run the five steps as separate passes, never as one prompt, with your rulebook as the authority on voice.
Filling in your own config: you do not invent your readers, your voice, or your rules from scratch. You research and mine them from material you already have, your own sent emails, past drafts, published pieces, and reactions to them, either by hand or with your agent's help. GUIDE.md has the step-by-step version with ready-made prompts; each file in config.template/ also links straight to its own step.
Method improvements are welcome as pull requests against SKILL.md, with the reasoning in the changelog. A useful test before opening one: would the rule hold for a team with a completely different voice and audience? If it depends on your voice, it belongs in your config, and we will say so in review. Voice never merges.
SHAPE is the portable part of the writing system at chim.ai, a Vienna-based AI transformation agency working with companies across the DACH region. chim.ai runs on a deliberately small human core and a board of AI agents that handle editorial work, sales operations and strategy synthesis. The method in this repository is the system that board writes with every week, and the corrections that trained its rulebook came from real published pieces, client engagements and a few documented failures. The weekly field notes from that setup appear in Running on Agents, chim.ai's LinkedIn newsletter about running a company on AI agents: what works, what breaks, what we declined.
chim.ai was founded by Dr. Andreas Fraunberger and Philipp Lammer. Andreas is an AI strategist who builds what he advises. His work sits at the seam of philosophy, creative direction and hands-on AI engineering: leadership programs and workshops for organizations adopting AI, augmentation consulting for teams putting agents into real operations, and Chimaira, a video-storytelling platform. The editorial judgment encoded in this method's rules is his; the method exists because those corrections kept repeating until they became a system. Philipp is chim.ai's CTO and creative developer, with a background in cybernetics; he owns the technical side of the company, including Chimaira and its Persona product.
The structural layer builds on StoryScope (Russell et al., 2026, arXiv:2604.03136, code and data): 10,272 prompts each written by a human and five LLMs, 61,608 stories, 93.2% macro-F1 separating human from AI on narrative structure alone, with per-model structural fingerprints. The method is model-aware because the fingerprints are: whichever model you draft on, gate hardest on its measured tell.
MIT. Use it, adapt it, build your own config on it. See LICENSE.