peithoSAFE

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: peitho
description: "Use when writing or revising any prose longer than a few sentences, in any form and any medium: article, blog post, marketing copy, landing page, README, email, chapter, story, social post, release notes, or the prose parts of a spec. Also use when a draft reads flat, generic, or AI-written, when an opening has to hook a reader who can leave instantly, when persuasion has to stay honest, or when asked to audit prose for slop. Also use when writing fiction, when a long piece needs a coherence pass, or when an ending has to close rather than trail off. Carries the em-dash law, the contraction default, the machine-prose tell catalog (hedging, process narration, restatement), the closing method, and the honest-persuasion rail. When the user supplies a draft they wrote, this skill's job is to edit it in their voice, never to recompose it."
---

# Peitho

Named for the Greek goddess of persuasion. This skill makes prose pull a reader in and hold them there, end to end, without lying to them.

It replaces `no-ai-slop`. That skill's content was sound; it simply never fired, because its trigger waited for the user to hand over a draft. This one fires whenever prose is being written.

## The axiom

**Pull is participation.** A reader stays because you have given them something to do. Three independent traditions, 2,300 years apart, converge on this:

- **Aristotle** built persuasion on the *enthymeme*: an argument that deliberately omits a premise so the audience supplies it.
- **Loewenstein (1994)** found curiosity is a *visible* gap. It dies at zero knowledge and it dies at complete knowledge. The reader must be able to see what is missing.
- **Heritage & Greatbatch (1986)** analyzed 476 real political speeches. Seven rhetorical formats accounted for about 70% of all applause. Contrast alone drew 33%. Every one of those formats opens a slot and lets the audience close it.

Two failure modes follow directly, and both are common:

- **Withhold everything** and there is no visible gap, so there is nothing to participate in.
- **Explain everything** and no slot is left, so there is nothing to participate in.

Decoration is not pull. Volume is not pull. Participation is pull.

## The three levels

Work all three. A perfect first line above flat paragraphs still loses the reader.

### Line level

- **Calibrate concreteness to an inverted U.** Too vague loses the reader; so does over-specification that closes the gap. A meta-analysis of 8,977 headline experiments found both tails underperform. Every paragraph needs at least one concrete anchor: a name, number, date, or mechanism.
- **Rhetorical figures are welcome.** Contrast, antithesis, tricolon, rhyme, chiasmus. McQuarrie & Mick (1996) found figures produced more elaboration and more favorable attitude *with no comprehension penalty*. Cleverness does not cost the reader.
- **Complexity is allowed.** Ashok, Feng & Choi (2013) classified successful novels at 84% accuracy and found they ran *more* syntactically complex, against readability. Correlational only, so do not chase complexity. But never flatten prose on the theory that short equals engaging.
- **Vary sentence shape.** Repeated shapes and stacked fragments read as machine output.
- **Make verbs carry the sentence.** "Decided" beats "made a decision."
- **Contract by default.** "It's," "don't," "you're," "we'll." Full forms are not more correct, they are a different register: contractions load on the *involved* pole of Biber's first dimension, and prose written entirely in full forms lands on the distant, informational pole whether the writer meant it to or not. It is also the most reliable surface tell of machine-written text. Expand only for stress, normative text, characterization, or clarity. Rule and its three abuses in `references/deslop.md`.

### Passage level

**The slot mechanic, and it runs the entire length of the piece:**

> Every paragraph closes one gap and opens another.

This is the hold engine. It rests on the Ovsiankina effect (open goals create a pull to resume), which replicates. It does **not** rest on the Zeigarnik memory claim (that unfinished things are better remembered), which a 2025 meta-analysis found does not replicate. Keep the motivational claim. Drop the memory claim.

Diagnostic for any paragraph after the first: **what did this close, and what did it open?** If neither, cut it or give it a job.

### Piece level

- **Serve all three exordium jobs at the door.** Classical rhetoric gives an opening three tasks, not one: make the reader *attentive*, *receptive*, and *well disposed*. One out of three is a failed opening.
- **Run a tension arc.** Boyd, Blackburn & Pennebaker (2020) analyzed roughly 40,000 narratives and found a consistent three-part shape: staging, then plot progression, then cognitive tension. Use it as structure. Do **not** believe it buys popularity: the same study found no relationship between structural adherence and how popular a story was. Structure is hygiene, not a growth lever.
- **Close every gap you opened.** By the end, inside the piece. This is both a craft rule and the first ethics rule.

## Entry module: openings

Openings are not a separate discipline. They are the three levels applied where exit risk peaks. Chartbeat, across two billion visits, found **55% of pageviews get under 15 seconds of active attention**, and the average reader quits around halfway.

Two inputs set everything:

1. **Exit speed.** How fast can this reader leave? A social reply is about one second. A blog lead is about fifteen. A chapter is minutes. Faster exit means the gap must be visible sooner and the concrete anchor must arrive earlier.
2. **Audience stance.** Receptive, neutral, or hostile. For a hostile or exhausted reader, classical rhetoric prescribes *insinuatio*, the oblique approach, rather than a frontal claim. Cicero named three causes of hostility, one being that the audience is simply worn out from listening. That

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (1)

og/source-artwork.png

Gates applied: no_behavioural_pass.