rephrasy

rephrasy-skills

Score text for AI detectability and humanize AI-generated text, without leaving Claude Code.

rephrasy-skills packages two Claude Code skills backed by the Rephrasy API: /rephrasy:ai-detect scores any text 0 (human) to 100 (AI) with a per-sentence breakdown, and /rephrasy:humanize rewrites AI-flavored text to read as human (Rephrasy's v3 model with the professional style by default). The skills encode live-verified API behavior — including the places where the API's real behavior diverges from its public docs.

Installation

Claude Code plugin

From inside Claude Code:

/plugin marketplace add jhubbardsf/claude-plugins
/plugin install rephrasy@joshd3v
/reload-plugins

joshd3v is Josh Hubbard's umbrella marketplace covering all of his plugins, so one marketplace add gives you this plugin and the rest.

Then run:

/rephrasy:ai-detect <file-or-text>
/rephrasy:humanize <file-or-text>

Direct repo install (alternative — this repo also embeds its own standalone marketplace):

/plugin marketplace add jhubbardsf/rephrasy-skills
/plugin install rephrasy@rephrasy-skills
/reload-plugins

Claude Code plugin install specs use plugin@marketplace. Here, rephrasy is the plugin name, joshd3v is the umbrella marketplace, and rephrasy-skills is this repo's standalone marketplace from .claude-plugin/marketplace.json. Plugin skills are namespaced by plugin name, so the invocations are identical either way.

One-line installer

curl -fsSL https://raw.githubusercontent.com/jhubbardsf/rephrasy-skills/main/install.sh | bash

To install at project or local scope:

curl -fsSL https://raw.githubusercontent.com/jhubbardsf/rephrasy-skills/main/install.sh | bash -s -- --scope project

Requirements

• Claude Code with plugin support
• python3 (the bundled helpers are stdlib-only — zero pip dependencies)
• A REPHRASY_API_KEY environment variable (get a key — this is a paid API; every call costs credits)

Usage

/rephrasy:ai-detect draft.md
/rephrasy:ai-detect "paste some prose here"
/rephrasy:humanize draft.md                       # -> draft_rephrasy.md next to the input
/rephrasy:humanize draft.md --verify              # re-score the result with the detector
/rephrasy:humanize draft.md --iterate --threshold 40
/rephrasy:humanize draft.md --style creative      # override the professional default
/rephrasy:humanize draft.md --model "Undetectable Model v2"

The humanize skill defaults to model v3 + style professional and never overwrites the input unless --in-place is passed explicitly. --iterate re-humanizes up to 3 passes, keeps the best-scoring version, and aborts (rather than burning paid passes) if the API errors.

Helper CLIs

The plugin bundles two stdlib-only Python scripts that also work standalone from any shell:

python3 plugins/rephrasy/scripts/rephrasy_detect.py FILE [--mode depth|default] [--json] [--threshold N] [--top K]
python3 plugins/rephrasy/scripts/rephrasy_humanize.py FILE [-o OUT] [--in-place] [--style STYLE] [--no-style] [--model MODEL] [--language NAME] [--json]

Both accept a file path, - for stdin, or --text "literal text".

Behavior	Detail
Exit codes	0 success · 1 over --threshold (detect only) · 2 any error — loops can distinguish "score too high" from "API is down"
Output naming	foo.txt → foo_rephrasy.txt (non-destructive by default); --in-place writes atomically via temp file + rename
Cost reporting	The humanizer prints cost=… credits to stderr after every call, including --json runs
Errors	Clean one-line error: … messages — no tracebacks, including response-phase timeouts and non-JSON 200 bodies

API facts verified live (2026-06-07)

These differ from, or go beyond, Rephrasy's public docs — the skills encode the observed behavior:

• Detector direction: 0 = human, 100 = AI in both default and depth mode. The docs claim non-depth modes are inverted; they are not.
• Default mode also returns per-sentence scores, just coarser than depth. The skills always use depth.
• The detector runs hot on formal prose: a verified-human (GPTZero 100%) formal sample scored 70.28; casual human text scored 7.35. The skills' verdict bands account for this.
• Humanizer costs is a plain number of credits, not the {"total": ...} object shown in the docs.
• new_flesch_score can be negative for dense prose.
• An invalid key returns HTTP 400 {"error":"Invalid API Key"}, not the documented 401 (both are handled).
• style: "professional" costs ~3x a no-style call and yields more formal (hence more Rephrasy-detectable) output. It stays the default because GPTZero — not Rephrasy's own detector — is the usual real target.

Configuration