Atlas / Skills / pr1m4lc0d3 / paraphrasis

paraphrasisSAFE

skills/pr1m4lc0d3/paraphrasis

Say the same thing to a different audience without the claim moving. A Claude Code skill: sense-for-sense restatement with a fidelity ledger and nine named ways a claim drifts under rewriting.

Verdict
SAFE
Grade
B
Trust score
89 /100
Version
—
Hosts
—
License
NOASSERTION
Stars
0
01

Overview

Say the same thing to a different audience without the claim moving. A Claude Code skill: sense-for-sense restatement with a fidelity ledger and nine named ways a claim drifts under rewriting.

Read from source at commit 81ad41edee0eOBSERVED · 2026-10-12
02

What it tells the agent

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: paraphrasis
description: Use when saying the same thing to a different audience — restating, re-voicing, translating for a new reader, shortening, expanding, or adapting a message for a developer, a buyer, an executive, a beginner, a skeptic, or a hostile reader. Also use when compressing to a word count, writing a summary or TL;DR, turning a spec into an announcement, a changelog into a release note, a paper into a post, or one channel's copy into another's. Carries the drift catalog, the fidelity ledger, and the reversal test. Not for improving prose in place — that is peitho. Not for rendering audio into text — that is metaphrasis.
---

# Paraphrasis

Sense-for-sense restatement, with the meaning held still.

Named against its opposite. **Metaphrase** is word-for-word; **paraphrase** is sense-for-sense. Dryden drew the line and put the danger on this side of it: once you stop moving words and start moving sense, nothing mechanical is left to check your work. This skill is the check.

## What this is for

One message, a different reader. A spec becomes an announcement. A finding becomes a landing page. Nine hundred words become ninety. The claim is supposed to survive all of it.

Usually it does not.

## The third job

`peitho` forks two ways: **Compose** when no draft exists, **Edit** when one does. Restating is neither, and running it as either causes damage.

| Job | Words | Meaning | Characteristic failure |
| --- | --- | --- | --- |
| Compose | new | new | slop, dishonest persuasion |
| Edit | mostly kept | kept | voice theft, recomposition |
| **Restate** | **replaced on purpose** | **must not move** | **drift** |

Edit protects the writer's voice by leaving the sentences alone. Restate cannot do that, because replacing the sentences is the entire assignment. So it needs a different guard, and the guard has to be on the meaning rather than the words.

## The axiom

**Everything may change except what is being claimed.**

| Free to change | Must survive unchanged |
| --- | --- |
| vocabulary, register, tone | claim strength |
| length, order, structure | scope and conditionality |
| examples, metaphors, framing | attribution and source |
| format, channel, medium | numbers and units |
| what to lead with | caveats and exclusions |
| how much context to supply | who or what is doing the acting |

If the right column moved, you did not paraphrase. You wrote a different claim in a similar voice, which is the failure this skill exists to catch.

## The fidelity ledger

**Every restatement returns one.** It is the deliverable, not a courtesy.

```
IN   →  Three seats on three providers returned three different rankings
        from one fixed prompt (Exhibit A, bb4aa857, 2026-08-15).

OUT  →  Ask three models the same question and you get three different
        answers. We logged it.

KEPT     three seats · three providers · three rankings · one prompt · logged
DROPPED  exhibit id, date          — recoverable from the source, not load-bearing here
ADDED    nothing
```

`ADDED` is the field that matters. Anything in the output with no ancestor in the input is invented, and inventing under a paraphrase reads as sourced when it is not. **A non-empty ADDED line is a stop, not a note.**

`DROPPED` needs a reason each time. "Cut for length" is only acceptable for detail that does not condition the claim. A caveat is never merely detail.

## The drift catalog

Named failures. Restatement finds these on its own; nobody has to intend them.

**Upgrade.** Hedged becomes absolute. *"seats disagreed in one session"* → *"councils reach better decisions."* The most common drift by a distance, because the upgraded version is always the more useful sentence.

**Scope creep.** A bounded instance becomes a general property. *"in one logged run"* → *"consistently."* *"on three of four providers"* → *"across providers."*

**Quantifier slide.** *some* → *most*. *up to* → *typically*. *can* → *does*. Each step is small and defensible; the chain is not.

**Agency shift.** *"can be configured to run locally"* → *"runs locally."* Capability restated as behavior.

**Evidence laundering.** The attribution drops, the claim survives, and the claim now floats. This is how "studies show" gets written by people who are not lying.

**Compression loss.** Under a word count the qualifier is the cheapest thing to cut and usually the load-bearing one. **When you shorten, cut examples before caveats.** If the caveat will not fit, the claim does not fit either.

**Register inflation.** Neutral → marketing → superlative, one restatement per step, each one only slightly warmer than the last.

**Nominalisation.** *"we measured X"* → *"X is proven."* The actor disappears and the finding hardens.

**Audience flattery.** Restating for a buyer quietly removes what the buyer would not want to hear. The audience may change what you lead with. It may not change what is true.

## The reversal test

Run before returning anything.

**Paraphrase the output back toward the input. Is the round trip stronger than the original?**

If restating your own OUT produces something the source would not support, you drifted, and the direction of the drift is now visible. Weaker is usually fine. Stronger never is.

## Hostile and skeptical readers

A skeptical reader is not a different tone. It is a different **burden of proof**, and it moves in one direction: toward more evidence and narrower claims, never fewer and broader. Copy for a skeptic that hedges less than the original has inverted the assignment.

For a hostile reader, state the strongest version of their objection before your answer. Peitho's ethics rail requires the counterargument anyway. Here it is also the only thing that gets read.

## The claims rail

**If the project has a claims file, check the output against it before returning.** Look for `truth.md`, `claims-register.md`, or a FORTRESS pack in a `.monkeys/` folder.

Restatement is the exact moment a barred claim re-ente
03

Trust audit

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.

LayerWhat it checksResult
L0Provenance & inventoryPASS
L1Static analysis of the codeNA
L2Instruction surface (what it tells the agent)PASS
L3Class-specific surfacePASS
L4Behavioural (sandbox)SKIPPED

What the source does

Filesystem
none-observed
Network
none-observed
Shell
none-observed
Dependencies
pinned
Secrets in source
none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.

Audited 2026-10-12 · audit v0.4.1 · source sha 81ad41edee0efull audit observations/trust-audit/skill/pr1m4lc0d3__paraphrasis.json · Report an issue / request a re-scan
04

Audit history

Every audit this skill has had.

DateSourceVerdictGradeScoreChange
2026-10-1281ad41edee0eSAFEB89first audit
05

Questions

What does the paraphrasis skill do?

Say the same thing to a different audience without the claim moving. A Claude Code skill: sense-for-sense restatement with a fidelity ledger and nine named ways a claim drifts under rewriting.

Is paraphrasis safe to install?

The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean skill reads B.

What can paraphrasis access on my machine?

The audit observed no filesystem, network or shell use at all in its source.

How current is this page?

The grade is for one exact copy of the source (81ad41edee0e), read on 2026-10-12. The repository is watched, and a new audit runs when it changes — this is the first audit.

Advertisement