A Claude Code plugin for writing articles, emails, letters, posts, and other communications that sound like you wrote them — not like AI.
Three slash commands, a per-piece context file, an audience-aware grading rubric, and a curated list of AI-writing tells to strip from every draft.
/scribe:draft → interview + first draft
/scribe:refine → interview + v2 (loops for v3, v4...)
/scribe:critique → 1-10 score + path to 10/10
Most AI writing has a fingerprint. Forced trios. "It's worth noting." "In conclusion." Mechanical bold labels. Em-dash overuse. Sycophantic openers. Reflexive both-sides framing. Detectors flag it; human readers can smell it even when they can't name it.
Scribe takes a different approach: it interviews you the way a ghostwriter would, then writes in your voice based on your own samples — with grammar fixes only when clarity needs them. Then it grades the result honestly and tells you exactly what to change to make it better.
Asks you, one or two questions at a time:
- Which saved voice to use (or create a new one from samples you paste)
- Who you are in this piece — role, stake, relationship to topic
- Who reads it — audience, formality, prior knowledge
- Format — email-to-boss / vendor / cold outreach / LinkedIn / op-ed / how-to / letter / memo / other
- Topic + concrete goal (what the reader should do, believe, or feel)
- Approach — argumentative / explanatory / story / warning / persuasive / informative
- Outline option — give bullet points first, or describe and let it structure
- Constraints — must-include facts, must-avoid words/topics, deadline
Writes the draft. Saves everything to .scribe-context.md in the current directory. Offers to save your voice to the catalog for future use.
Reads the context file. Asks:
- Read it again. What landed? What felt off?
- Opener strong, or throat-clearing?
- Ending lands, or trails?
- Any sentence that doesn't sound like you?
- Anything to add? Cut? Reframe?
- Did your goal shift while reading?
Produces v2. Appends to context file. Loops indefinitely for v3, v4.
Picks the rubric matching your format (LinkedIn post is weighted differently than an escalation email or an op-ed).
Outputs:
- Overall: X.X / 10, with weighted dimension scores
- Top 3 issues, each with exact location + concrete fix
- What's working (so the next refine doesn't quietly delete it)
- Path to 10/10 — for every below-10 dimension, an exact paragraph + exact edit
- Top 3 highest-leverage moves ranked by score-impact-per-effort
- Stretch moves specific to the format (op-eds need one quotable line; LinkedIn needs a number in the hook; etc.)
- Ceiling note when the format honestly caps the achievable score
No grade inflation. No vague advice. Every below-10 score comes with a concrete edit.
Scribe is a Claude Code plugin. This repo is both the plugin and its own marketplace, so installing it is two commands inside Claude Code:
/plugin marketplace add BuggyButLearning/scribe
/plugin install scribe@buggybutlearning
The first command registers this repo as a marketplace (named buggybutlearning); the second installs the scribe plugin from it. Reload when prompted.
Verify: type /scribe: in any project and you should see draft, refine, and critique.
Works on every project automatically. No per-project setup.
/plugin marketplace update buggybutlearning
- Saved voice profiles are written to
~/.claude/scribe/voices/— your home directory, not the plugin cache — so they survive plugin updates and are shared across every project. - The per-piece context file (
.scribe-context.md) is written to whatever directory you run the commands in, so each piece stays with its project.
cd <wherever you want the context file saved>
/scribe:draft
Scribe asks for samples (paste them, or give a file path). Answers a few questions. Writes the draft. Offers to save your voice for next time.
/scribe:refine
Reads .scribe-context.md. Asks what worked and what didn't. Writes v2. Repeat as many times as you want.
/scribe:critique
Reads the latest version. Scores it. Tells you exactly what would push it to 10.
The context file lives in your project folder. Close the session, come back tomorrow, run /scribe:refine — it picks up where you left off.
Use .scribe-context-<slug>.md to keep contexts separate. Scribe will prompt if it finds an existing context.
Saved profiles live at ~/.claude/scribe/voices/<slug>.md. Each one captures:
- Sentence-length patterns and rhythm
- Vocabulary level, signature words, jargon comfort
- Habits (contractions, em-dashes, paragraph openers, punctuation tics)
- Things to avoid (your preferences)
- Source sample excerpts for future reference
Build a library: professional-engineer, dad-to-school, linkedin-personal, cold-outreach-warm, whatever fits how you write across contexts. /scribe:draft lists them and lets you pick per piece.
A short slice of what's in resources/ai-tells.md:
- Vocab: delve, leverage, robust, pivotal, comprehensive, seamless, tapestry, navigate (figurative), nuanced, intricate, stands as, serves as, underscores...
- Phrases: "I hope this email finds you well," "just checking in," "in today's landscape," "in conclusion," "let's dive in," "it's worth noting," "from X to Y..."
- Structural: forced three-item lists, mechanical bold labels, uniform paragraph length, em-dash overuse, smooth transitions on every paragraph, reflexive both-sides framing
- Tone: sycophancy, fake warmth, over-apology, knowledge-cutoff disclaimers, closing pep talks
Every draft and revision runs a final "what about this still sounds AI?" pass.
resources/voice-rubric.md covers:
Universal dimensions (always graded, 50% total):
- Clarity, Voice match, Audience fit, Structure, Human-ness
Format-specific dimensions (50% total): one block per format —
- Email to boss · Email to vendor · Cold outreach · Internal team update · Difficult/sensitive communication · LinkedIn post · Blog post / article · Op-ed · How-to / explainer · Personal letter · Memo / report
Each dimension has anchor descriptions for scores 1, 4, 7, 10 so the grading stays honest and repeatable across pieces.
scribe/ ← this repo = the plugin (and its own marketplace)
├── .claude-plugin/
│ ├── plugin.json plugin manifest (name: scribe)
│ └── marketplace.json marketplace manifest (name: buggybutlearning)
├── commands/
│ ├── draft.md /scribe:draft
│ ├── refine.md /scribe:refine
│ └── critique.md /scribe:critique
├── resources/
│ ├── ai-tells.md AI-writing tells to strip
│ ├── voice-rubric.md 1-10 grading rubric, format-aware
│ └── interview-questions.md grill-me question banks for draft + refine
├── voices/
│ └── index.md seed catalog (your saved voices live in ~/.claude/)
├── assets/
│ ├── scribe-banner.png the title image
│ └── scribe-banner.spec.json the JSON spec it was generated from
├── README.md
└── LICENSE
# Created on your machine as you use it:
~/.claude/scribe/voices/ your saved voice profiles (persist across updates)
~/.claude/scribe/voices/<slug>.md one file per saved voice
.scribe-context.md per-piece context + all revisions (in your working dir)
The command files reference their bundled resources with ${CLAUDE_PLUGIN_ROOT} so they resolve wherever the plugin is installed.
/scribe:draft
↓
context file + v1
↓
/scribe:refine ←┐
↓ │
v2, v3, v4... │ loop until you're happy
↓ │
/scribe:critique │
↓ │
score + path to 10
↓ │
if < 8: refine ─┘
if ≥ 8: ship
- Sound like a person, not an AI. Voice match is weighted equally to clarity. Detector-tripping fingerprints are explicitly enumerated and removed.
- Grammar fixes only when clarity needs them. If your samples end sentences with prepositions or open paragraphs with "And," scribe keeps that.
- Interview, don't assume. Five-question minimum before drafting. Skip nothing important. Don't pile every question into one wall.
- Honest grading. Every score below 10 comes with an exact edit. No vague advice. Ceiling notes when format genuinely caps the achievable score.
- One piece, one context file. Survives session breaks. Refine days later without re-explaining anything.
- Composable with itself. Three commands, one workflow. Same pattern as
/paul— each command suggests the next.
MIT — do what you want. Pull requests welcome on the rubric, tells list, and format-specific dimensions.