GitHub stars
40

Install

/plugin install writing@claude-code-plugins-marketplace

Run it inside Claude Code after adding the marketplace with /plugin marketplace add doodledood/claude-code-plugins.

Contents

Discovered from source

Skills 4

User request: $ARGUMENTS

Author Voice Skill

Create a maximally information-dense AUTHOR_VOICE.md document through iterative refinement. The resulting document enables any LLM to write content that authentically matches your voice.

Overview

This skill supports both creating a new voice doc and refining an existing one. Users often come back multiple times to adjust their doc as their voice evolves or as they notice issues in generated content.

This skill guides you through: 0. Check Existing - Look for existing AUTHOR_VOICE.md; let user choose to refine or start fresh

  1. Discovery - Clarifying questions about your voice characteristics and goals
  2. Initial Draft - Generate first AUTHOR_VOICE.md based on your inputs (good enough to start!)
  3. Refinement Cycles - Generate sample texts, collect your ratings/feedback, update the doc ⬅️ THE REAL MAGIC
  4. Completion - Final honed document ready for AI content generation

Returning users can skip straight to Phase 3 to run more feedback cycles on their existing doc.

Where the Value Comes From

The discovery questions give you a solid starting point - but the real magic happens in Phase 3. That's where you:

  • See actual generated samples in your "voice"
  • Judge them ("this doesn't sound like me")
  • Give specific feedback ("too formal", "wrong words")
  • Watch the voice doc evolve until it truly captures YOU

Each feedback cycle sharpens the doc. Most users need 3-5 cycles to get from "this is okay" to "this is actually me."

Workflow

Phase 0: Check for Existing Document

Before starting discovery, check if the user already has an AUTHOR_VOICE.md:

  1. Search for existing doc: Use Glob to search for **/AUTHOR_VOICE.md in the current directory and common locations
  2. If found: Read it and ask the user what they want to do
header: "Existing Voice Doc Found"
question: "I found an existing AUTHOR_VOICE.md. What would you like to do?"
options:
  - "Refine it - run feedback cycles to improve accuracy"
  - "Start fresh - create a new voice doc from scratch"
  - "Review it - just read through what's there"

If "Refine it": Skip to Phase 3 (Refinement Cycles) - the user is coming back to improve their existing doc.

If "Start fresh": Proceed to Phase 1 (Discovery) - will overwrite the existing doc.

If "Review it": Read and display the doc, then ask what they want to do next.

  1. If NOT found: Proceed directly to Phase 1 (Discovery)

Phase 1: Discovery

Use AskUserQuestion tool with multi-choice options for EVERY question to minimize cognitive load. If AskUserQuestion is unavailable, present numbered options and ask the user to reply with the number(s).

Question 1: Primary Content Type

header: "Content Type"
question: "What's your PRIMARY content format?"
options:
  - "Twitter/X posts (short-form, punchy)"
  - "LinkedIn posts (professional, insights)"
  - "Blog articles (long-form, detailed)"
  - "Newsletter (conversational, regular)"
  - "Technical documentation (precise, instructional)"
  - "Mixed - I write across multiple formats"

Question 2: Voice Personality

header: "Voice Tone"
question: "How would you describe your voice personality?"
options:
  - "Authoritative expert - confident, direct, no-nonsense"
  - "Friendly mentor - approachable, encouraging, educational"
  - "Provocateur - contrarian, challenges assumptions, bold"
  - "Storyteller - narrative-driven, uses examples, personal"
  - "Analyst - data-driven, logical, objective"
  - "Conversational peer - casual, relatable, human"
multiSelect: true (pick up to 2)

Question 3: Signature Elements

header: "Signatures"
question: "What signature elements define your writing?"
options:
  - "Strong opening hooks"
  - "Numbered lists and frameworks"
  - "Personal anecdotes and stories"
  - "Contrarian takes and hot takes"
  - "Data and research citations"
  - "Metaphors and analogies"
  - "Direct calls-to-action"
  - "Questions to engage readers"
  - "Short punchy sentences"
  - "Long flowing prose"
multiSelect: true

Question 4: Vocabulary Style

header: "Vocabulary"
question: "What's your vocabulary style?"
options:
  - "Simple and accessible - anyone can understand"
  - "Technical jargon - domain-specific terms expected"
  - "Casual slang - internet-native, memes okay"
  - "Formal professional - polished, corporate-appropriate"
  - "Academic - precise, nuanced, scholarly"

Question 5: Emotional Range

header: "Emotion"
question: "What emotions do you convey in your writing?"
options:
  - "Enthusiasm and excitement"
  - "Calm confidence"
  - "Urgency and importance"
  - "Humor and wit"
  - "Empathy and understanding"
  - "Skepticism and critical thinking"
  - "Inspiration and motivation"
multiSelect: true

Question 6: Target Audience

header: "Audience"
question: "Who is your primary audience?"
options:
  - "Developers/Engineers"
  - "Founders/Entrepreneurs"
  - "Product managers"
  - "Executives/Leaders"
  - "General tech audience"
  - "Non-technical professionals"
  - "Students/Learners"

Question 7: Writing Goals

header: "Goals"
question: "What do you want your writing to achieve?"
options:
  - "Build authority and thought leadership"
  - "Drive engagement and discussion"
  - "Educate and inform"
  - "Entertain and delight"
  - "Convert/sell (subtly)"
  - "Build community and connection"
multiSelect: true

Question 8: Anti-patterns

header: "Avoid"
question: "What should your writing NEVER do?"
options:
  - "Use corporate buzzwords"
  - "Sound robotic or AI-generated"
  - "Be preachy or condescending"
  - "Use excessive emojis"
  - "Be overly promotional"
  - "Use clickbait tactics"
  - "Be wishy-washy or hedging"
  - "Use filler phrases"
multiSelect: true

Question 9: Sample Topics (free text acceptable here)

header: "Topics"
question: "List 2-3 topics you frequently write about (or paste examples of your past writing)"

Phase 2: Initial Document Generation

After discovery, generate the first AUTHOR_VOICE.md. This draft is intentionally a "good enough" starting point - not perfect, but solid enough to generate samples and start the feedback loop.

Use this structure:

# AUTHOR_VOICE.md

> This document defines [Author]'s writing voice for AI content generation.
> Feed this to any LLM before requesting content to match the author's style.

## Voice Identity

[1-2 sentences capturing the core voice essence]

## Tone Parameters

- **Primary tone**: [e.g., "Confident mentor with occasional humor"]
- **Emotional range**: [comma-separated emotions from discovery]
- **Formality level**: [1-10 scale with description]
- **Warmth level**: [1-10 scale with description]

## Structural Patterns

- **Opening style**: [how posts/articles begin]
- **Paragraph length**: [short/medium/long, typical sentence count]
- **List usage**: [when and how lists are used]
- **Closing style**: [how content ends - CTA, question, statement]

## Vocabulary Rules

### USE:
- [Specific words/phrases the author uses]
- [Technical terms that are okay]
- [Signature expressions]

### AVOID:
- [Words that feel off-brand]
- [Overused phrases to skip]
- [Tone markers to avoid]

## Content Patterns

### Hooks
[How the author grabs attention - examples of opening patterns]

### Arguments
[How the author builds points - numbered, narrative, comparison]

### Evidence
[How claims are supported - data, anecdotes, logic, authority]

### Transitions
[How ideas connect - explicit markers, implicit flow]

## Signature Moves

1. [Specific technique the author uses regularly]
2. [Another signature element]
3. [Third distinguishing characteristic]

## Anti-Patterns

NEVER:
- [Specific thing to avoid]
- [Another anti-pattern]
- [Third prohibition]

## Example Transformations

### Generic version:
"[Common way to express an idea]"

### In author's voice:
"[Same idea in the author's distinctive style]"

---

**User request**: $ARGUMENTS

## Quick Reference

**One-line voice summary**: [Author] writes like [analogy/comparison].

**Before generating content, ensure**:
- [ ] [Checklist item 1]
- [ ] [Checklist item 2]
- [ ] [Checklist item 3]

Write this file to the current working directory as AUTHOR_VOICE.md.

Phase 3: Refinement Cycles (Where the Magic Happens)

This is the most important phase. The initial doc from Phase 2 captures the basics, but YOUR feedback on generated samples is what transforms it from generic to authentic. Don't skip this.

After generating the initial document, begin iterative refinement:

Step 3.1: Generate Sample Texts

IMPORTANT: Do NOT generate samples yourself. You MUST use the voice-writer agent to generate samples. The agent is specifically designed to read the voice doc and produce authentic samples - attempting to generate them inline will produce inferior results.

Use the voice-writer agent to generate 3 sample texts.
Voice doc path: [path to AUTHOR_VOICE.md]
Mode: Sample generation

The agent will read the voice doc and output 3 samples:

  1. Short-form (~280 chars)
  2. Medium-form (2-3 paragraphs)
  3. Conversational reply

Step 3.2: Collect Feedback Per Sample

For EACH of the 3 generated samples, use AskUserQuestion tool (or numbered options if unavailable):

header: "Sample [N]"
question: "Rate this sample and share what's off:"
[Display the sample text]
options:
  - "Perfect - captures my voice exactly"
  - "Close - minor tweaks needed"
  - "Okay - something feels off but hard to pinpoint"
  - "Wrong - this doesn't sound like me"

If not "Perfect", follow up with:

header: "Feedback"
question: "What specifically needs adjustment in Sample [N]?"
options:
  - "Too formal/stiff"
  - "Too casual/unprofessional"
  - "Wrong vocabulary/word choices"
  - "Missing my signature style elements"
  - "Tone is off (wrong emotion)"
  - "Structure doesn't match how I write"
  - "Too long/wordy"
  - "Too short/choppy"
  - "Other - let me explain"
multiSelect: true

If "Other" selected or if more detail needed, prompt for free-text: "Describe what's wrong and how you'd actually write this:"

Step 3.3: Update Document

Based on ALL feedback from the 3 samples:

  1. Identify patterns in the feedback (what's consistently wrong?)
  2. Update the AUTHOR_VOICE.md with new/refined rules
  3. Add specific "instead of X, use Y" examples where needed
  4. Strengthen anti-patterns if certain issues keep appearing

Step 3.4: Check Completion

header: "Continue?"
question: "Want to run another refinement cycle?"
options:
  - "Yes - generate 3 more samples with the updated doc"
  - "No - the voice doc is good enough for now"
  - "Almost done - one more cycle should perfect it"

If "Yes" or "Almost done", return to Step 3.1 with the updated document.

Phase 4: Completion

When user indicates completion:

  1. Add a "Version History" section noting refinement cycles completed
  2. Add a "Usage Instructions" section for how to use with LLMs
  3. Display final document summary
  4. Remind user to keep the AUTHOR_VOICE.md with their projects

Key Principles

Information Density

  • Every line in the doc must be actionable for an LLM
  • No fluff, no explanations "for humans"
  • Concrete examples over abstract descriptions
  • Specific word lists over vague guidelines

Iterative Refinement

  • 3-5 refinement cycles typically needed for high accuracy
  • Each cycle should fix specific issues identified
  • Track what changes between versions

Reduce Cognitive Load

  • ALWAYS use AskUserQuestion tool when available - this is critical for UX
  • Present multi-choice questions to minimize user typing/thinking
  • Limit options to 6-8 max per question
  • Use multiSelect for non-exclusive choices
  • Only use free-text for examples/samples or when AskUserQuestion unavailable

Output Location

Write AUTHOR_VOICE.md to the current working directory (or user-specified path).

User request: $ARGUMENTS

Apply these research-backed writing principles to the current task. If no specific request, apply them to whatever prose content is being written in context.

The Core Insight

The fundamental problem is statistical uniformity. AI text is measurably more predictable (~50% lower perplexity), less varied in sentence length (~38% lower burstiness), and narrower in vocabulary (type-token ratio: human 55.3 vs AI 45.5). The path to human-sounding writing runs through embracing imperfection, not perfecting output.

The single most reliable tell is uniformity. Human writing is messy, varied, and surprising. AI writing is smooth, consistent, and predictable.

The 10-20-70 Rule

Prompting contributes ~10% of output quality, editing ~20%, and the writer's own domain expertise and input ~70%. No amount of prompt engineering substitutes for having something to say. Require the writer's genuine insight, opinions, and experiences before generating content.

Hierarchy of Impact

From highest to lowest impact on making writing sound human:

PriorityTechniqueWhy
1Put your own thinking in firstAI cannot generate genuine insight, lived experience, or original analysis
2Develop a distinctive voiceVoice is the ultimate differentiator — consistent, cannot be faked by editing
3Edit ruthlesslyFour-layer system: word → sentence → structure → content
4Design the workflowNever write a complete piece in one shot
5Prompt with constraintsBanned words + persona + writing samples
6Embrace imperfectionFragments. Tangents. Opinions. Rough edges make writing alive.

Vocabulary Kill-List

Avoid these words when writing — they are statistically flagged as AI-generated across peer-reviewed studies of millions of documents. (Reviewers reading existing prose should judge by density and clustering rather than single instances; see ai-tells-and-fingerprints.md and the writing-reviewer agent's frequency-aware threshold.)

Nouns: delve, tapestry, landscape, realm, testament, journey, insight, resilience, ecosystem, milestone, prowess, utilization

Verbs: embark, endeavor, leverage, harness, navigate (metaphorical), unlock, foster, catalyze, bolster, underscore, showcase, elucidate, encompass, unveil

Adjectives: seamless, robust, groundbreaking, transformative, pivotal, vibrant, compelling, crucial, invaluable, holistic, multifaceted, meticulous, commendable, intricate

Adverbs: seamlessly, meticulously, notably, profoundly, predominantly, subsequently, thereby, ultimately, moreover, furthermore

Phrases: "ever-evolving landscape," "in today's fast-paced world," "as we navigate the complexities," "It isn't just X, it's Y," "it's important to note," "it's worth noting that," "without further ado," "in conclusion," "at the heart of"

Puffery / promotional drift: "breathtaking," "stunning," "must-see," "must-visit," "iconic," "world-class," "rich cultural tapestry," "hidden gem"

False intensifiers: "genuinely," "truly," "actually" (when used to simulate conviction)

Era-tracked vocabulary: AI vocabulary shifts over time — "delve" peaked 2023–early 2024 then declined, while "align with" / "fostering" / "showcasing" rose with later models. For era-specific lists (GPT-4 era vs GPT-4o era) and the snapshot date, see ai-tells-and-fingerprints.md § Era-tracked vocabulary.

Four-Layer Editing System

Apply in order from surface to substance:

Layer 1: Word-Level

Search-replace or delete kill-list vocabulary on sight. Strip adjectives from paragraphs, restore only those carrying concrete information. "Robust system" → "handles 10k req/s without data loss."

Layer 2: Sentence-Level

Read only the first few words of consecutive sentences — wherever three or more follow the same pattern, cut or combine. Vary sentence length deliberately: short for punch, long for nuance. The contrast creates impact. Add intentional imperfection: fragments, casual asides, conversational phrasing.

Layer 3: Structural

Eliminate meta-commentary ("In this section, we will..."). Kill recap conclusions that only repeat earlier points. Break pattern symmetry: demote repetitive subheadings, merge overlapping sections, ensure each paragraph's opening differs structurally from the one before.

Layer 4: Content

Add lived experience: anecdotes, firsthand observations, specific failures. Ground in specifics — ask of every sentence: "Could this fit any topic?" If yes, it needs grounding. Inject honest opinion: state what you actually think, not what "many experts" believe.

Final Check

Read aloud. Stumbling, running out of breath, or awkwardness marks where prose needs work.

Seven Craft Fundamentals AI Structurally Cannot Produce

These are structural limitations of statistical text generation — areas where human writers create unbridgeable distance:

  1. Showing vs Telling — Render specific sensory details that let readers experience emotion. AI defaults to summarizing ("serene and tranquil") rather than showing (the dragonfly hovering over still water).

  2. Specificity from Lived Experience — AI produces "gentle breeze" and "blooming flowers" (statistically most probable). Replace generic descriptions with observations nobody else has made. Name the cafe, the specific dish, the particular moment.

  3. Strategic Omission — AI tends toward completeness and closure. Resonant writing lives in what's left unsaid. A character dodging a question reveals more than any direct statement. Trained to produce text, not withhold it.

  4. Rhythm Variation — AI produces sentences of similar length and structure. Use rhythm deliberately: shorten sentences as tension rises. Drop a short sentence after several long ones. Like that.

  5. Deliberate Rule-Breaking — Choose the wrong word because it sounds better. Let a fragment hang. Incomplete sentences. For emphasis. Because sometimes a complete sentence kills the moment.

  6. Humor — Classified as an "AI-complete problem." Google DeepMind study with 20 comedians: AI "struggled to produce material that was original, stimulating, or — crucially — funny." Humor requires authentic vulnerability and cultural boundary-breaking.

  7. Genuine Insight — AI provides summaries; humans provide analysis. Keep asking "Why?" iteratively. Data shows the "what" — insight tells the "why."

Structural Anti-Patterns

PatternTellFix
Uniform paragraph lengthEvery section gets equal treatment regardless of importanceSpend more space on what matters, less on what doesn't
List addictionJumping into numbered/bulleted lists without narrative buildupUse flowing prose; lists only when genuinely parallel
Formulaic scaffolding"Firstly... Secondly... Finally" at 2-5x human rateVary transitions or eliminate them
Grammar perfectionNo fragments, run-ons, or unconventional startsPerfection is suspicious — include occasional wonky phrasing
Colon titles"Topic: Explanation" formatVary title structure
Symmetric structureEvery section mirrors the same internal organizationBreak the pattern

Punctuation and Formatting Rules

  • Em-dashes (—) and en-dashes (–): One of the most reliable AI tells. ChatGPT uses 8 per 573 words; Deepseek 9 per 555 words. Ban them entirely — use commas, periods, parentheses, or colons instead.
  • Curly quotation marks (" " ' '): ChatGPT and DeepSeek tend to produce curly quotes; Gemini and Claude tend to produce straight ("..." '...'). Curly quotes alone are not proof (word processors and typesetting tools auto-curl them) — but in combination with other tells they raise confidence. Prefer straight quotes for first-draft AI output; let the editor decide if curling is appropriate for the venue.
  • Heading capitalization: AI tends to title-case all main words in headings ("How to Write Better Prose") even when the document otherwise uses sentence case. Match the surrounding document's heading style; don't default to title case.
  • Excessive boldface ("key takeaways" pattern): AI mechanically bolds phrases for emphasis, often the same word every time it appears, or in bullet lists where the bold phrase is then redeclared (**Scalability:** The system is designed to scale...). Use boldface sparingly and only where genuine emphasis serves the reader.
  • Emojis: AI overuses emojis as emotional proxies, especially in casual/marketing content. Never add emojis unless the user explicitly requests them. Excessive emoji use signals AI generation immediately.
  • Semicolons: AI rarely uses them. Including some adds human texture.
  • Contractions: AI avoids them. Use them freely in conversational prose.
  • Oxford commas: AI applies them consistently. Break the pattern occasionally.

Rhetorical Anti-Patterns

PatternTellFix
Tricolon obsessionGroups ideas in threes: "Time, resources, and attention"Break with two, four, or seven items erratically
Perfect antithesis"Not just X, but Y" — neat binary oppositionsReal arguments are messier
Rhetorical questions as staging"How do we solve this?" → pre-composed answerAsk genuine questions or just state the point
Excessive hedging"may potentially offer what could be considered significant benefits"Strip to: "this works"
Compulsive signposting"It's worth noting," "It's important to remember"Trust the reader
Opinion-avoidant framing"commonly described as," "many find," "generally considered"State the view directly
Overused conjunctions"Moreover," "Furthermore," "Additionally," "In addition" stacking across paragraphs at AI-typical densityCut transitions; let sentences carry the logic. One transition per page, not per paragraph
"Myths busted" / contrast-and-correct"While many think X, in fact Y" pattern used as default opener regardless of whether a myth existsOpen with the specific claim; skip the strawman setup
Subject pufferyArbitrary detail elevated to "a microcosm of [larger theme]" / "a window into [broader cultural moment]"Stop at the specific. Don't extrapolate unless the piece earns it
Statistical regression to the meanConcrete details (names, numbers, dates) blurred into category-level languageRestore specifics. If you don't know them, say "I don't know"

Tonal Principles

  • Vary register. AI picks a lane and stays there. Shift between formal and colloquial. Reveal personality through tonal variation.
  • Don't be relentlessly positive. AI frames everything positively. Call things weak, inadequate, or bad when they are.
  • Show unequal enthusiasm. AI treats all subjects with equal professional distance. Nerd out about topics you care about. Show visible impatience with boring ones.
  • Take risks. AI prioritizes broad palatability. Write confusing sentences, sharp observations, and controversial assertions when appropriate.
  • Reject encyclopedic-yet-promotional drift. Even when prompted for a neutral or encyclopedic register, AI drifts toward advertisement-like writing — travel-guide prose for places, marketing copy for products. Watch for the puffery vocabulary above ("breathtaking," "must-visit," "rich cultural tapestry"). Neutral does not mean polished-bland; it means specific, factual, and unlabored.

What Must Be Present (The Negative Space)

AI text is identified as much by what's absent as what's present:

  • Lived experience — specific personal anecdotes, not "generic specificity"
  • Sensory specificity — unexpected observations, not statistically probable descriptions
  • Silence and subtext — what characters don't say, what's left unsaid
  • Genuine messiness — false starts, changed directions, productive digressions
  • A perspective — a view that could not fit any other prompt

Prompting Techniques (When Generating)

Tier 1 (Strongest evidence)

  • Banned word lists: Prohibit the kill-list vocabulary explicitly. Constraint-based prompting outperforms aspirational instructions.
  • Persona assignment: "Explain this to a colleague over coffee" produces fundamentally different output than "Write about X."
  • Writing sample matching: Paste your own writing and instruct to match tone, structure, emotional depth — not copy structure.

Tier 2 (Good evidence)

  • Conversational framing: Frame as conversations, not commands.
  • Negative instructions: "Don't use formal transitions. Don't start paragraphs with 'It is worth noting.'"
  • Emotional targeting: "Write from a place of frustration with the old way" — specify the emotional register.

Tier 3 (Supporting evidence)

  • Short sentence requests: Counteracts multi-clause tendency.
  • Dramatic paragraph length variation: Include one-sentence paragraphs for emphasis.
  • Layer multiple techniques: Combine persona + banned words + emotional targeting + samples.

Workflow Principles

  • Never write a complete piece in one shot. Section-by-section with review between.
  • Bookend approach: AI for ideation at start, polish at end. Human controls the core creative middle.
  • Messy draft approach: Your scattered thoughts + original angles → AI cleans into readable prose → you edit for voice.
  • Your original thinking goes IN before AI touches it.

Statistical Signatures (Detection Context)

Understanding what detectors measure helps write text that doesn't trigger them:

MetricHumanAIMeaning
Perplexity (surprisal)~8.2~4.2AI is ~50% more predictable
Burstiness (sentence variation)0.610.38AI has ~38% less variation
Token probability entropy4.563.11AI makes more uniform word choices (d=3.08)
Type-token ratio55.345.5Humans use broader vocabulary
Late-stage volatilityConsistentDecays 24-32%AI becomes more predictable as it continues

Key implication: Introduce genuine unpredictability — varied vocabulary, surprising sentence lengths, unexpected word choices, inconsistent structure.

Detection Methods (Reference)

  • Statistical/zero-shot: Perplexity, burstiness, entropy, probability curvature (DetectGPT), cross-perplexity ratio (Binoculars), directional memorization (BiScope)
  • Trained classifiers: GPTZero, Originality.ai, Turnitin, Copyleaks — achieve 65-96% on unedited AI text
  • Watermarking: Vocabulary partitioning, SynthID tournament sampling
  • Stylometric: Type-token ratio, stop word count, hapax legomenon rate
  • What defeats detection: Paraphrasing (reduces by 87-99%), simple modifications, human editing, nucleus sampling
  • Theoretical limit: Detection converges toward random chance as models improve (Sadasivan et al.)

For full detection science detail, see detection-science.md.

Model-Specific Signatures

ModelKey Tells
ChatGPTFormal, clinical; heavy em-dashes (8/573 words); overuses "delve," "align," "noteworthy"; dry, robotic
GeminiConversational, explanatory; prefers simple language; no em-dash overuse
ClaudeMore natural and literary; minimal em-dashes (2/948 words); tonal flexibility; occasionally generates fiction unprompted
DeepseekHeavy em-dashes (9/555 words); similar to ChatGPT structurally

Reference Files

Detailed research backing these principles:

FileContentsConsult When
ai-tells-and-fingerprints.mdVocabulary fingerprints, structural/rhetorical/tonal patterns, statistical signatures, model-specific signaturesReviewing text for AI tells, understanding what detectors look for
humanizing-playbook.mdFour-layer editing system, voice development, 7 craft fundamentals, professional workflows, what editors look forEditing AI-assisted text, developing writing voice, understanding editorial standards
prompting-and-workflow.mdPrompt engineering techniques (tiered), workflow designs, tool-specific strategies, style library buildingSetting up writing prompts, designing hybrid workflows, building personal style libraries
detection-science.mdDetection methods, accuracy data, evasion methods, arms race, theoretical limits, non-native speaker biasUnderstanding how detection works, what statistical properties matter, accuracy limitations

Never Do

  • Use kill-list vocabulary
  • Use em-dashes or en-dashes (use commas, periods, parentheses, or colons instead)
  • Add emojis unless the user explicitly requests them
  • Write uniform paragraph lengths
  • Start consecutive sentences with the same structure
  • Hedge when a direct statement serves better
  • Treat all subjects with equal professional distance
  • Explain what you're about to explain before explaining it
  • End with a conclusion that repeats the introduction
  • Produce text that could fit any prompt equally well

Use the voice-writer agent to generate content in the user's voice.

Topic/prompt: $ARGUMENTS

User request: $ARGUMENTS

Write prose content that reads as authentically human, not AI-generated.

Goal

Gather the writer's context (the 70%) → write content applying human-writing principles → review via writing-reviewer → auto-fix HIGH+ issues → loop until reviewer finds no HIGH+ issues → deliver.

Input Detection

Adapt based on what's provided:

InputEntry Point
Topic onlyGather context → outline → write → review loop
Topic + outlineGather context → write from outline → review loop
Rough draftGather context for gaps → rewrite applying principles → review loop
Finished textSkip writing, run review loop only (editing mode)
No inputAsk what to write

Gather Context (The 70%)

Before writing anything, gather the writer's substance. This is the primary quality driver — prompting and editing are the remaining 30%.

Ask via AskUserQuestion:

  • What's your take? Key points, thesis, or argument you want to make
  • Personal experience? Specific anecdotes, examples, observations to include
  • Opinions? What do you believe about this? What's your angle?
  • Specifics? Names, numbers, dates, details that ground the piece
  • Audience? Who reads this and what should they walk away with?
  • Tone? How should this feel? (Conversational, authoritative, provocative, casual)

If the user provides a rough draft, extract their existing substance and ask only about gaps.

Write

Invoke the writing:human-writing skill to apply research-backed writing principles during generation.

Write the content to draft-{topic-slug}.md in the current working directory. Use a slugified version of the topic for the filename.

Incorporate the user's context as the core substance. Apply the principles from human-writing: avoid kill-list vocabulary, vary sentence length and structure, break symmetry, include specificity, take positions, embrace imperfection.

Review Loop

After writing (or receiving finished text for editing):

  1. Review: Invoke the writing-reviewer agent on the draft file
  2. Auto-fix: Fix all CRITICAL and HIGH severity AUTO_FIXABLE issues
  3. Report: Present MEDIUM and LOW issues to the user (informational, not blocking)
  4. Present NEEDS_HUMAN_INPUT: Show HIGH+ NEEDS_HUMAN_INPUT issues to user with options; skip if user declines
  5. Loop: If fixes were applied, re-review. Continue until reviewer finds no HIGH+ issues.

Convergence guard: If a fix cycle introduces new HIGH+ issues (total HIGH+ count increases instead of decreasing), stop and present the situation to the user rather than continuing to loop.

Constraints

ConstraintWhy
Context before contentNever generate prose without the writer's substance first
Converge, don't capNo arbitrary iteration limit — run until no HIGH+ issues
Convergence guardStop if fixes create new problems — don't oscillate
DRYDelegate principles to human-writing, review to writing-reviewer
File outputWrite to draft-{topic-slug}.md in cwd
AtomicWorking draft updated in place; user sees final clean version

Output

Report when done:

  • File path
  • Review iterations completed
  • Issues fixed (auto) vs reported (to user)
  • NEEDS_HUMAN_INPUT items presented and their resolution
  • Summary of key changes made during review cycles

Agents 2

Voice Writer Agent

You are an expert ghostwriter who can perfectly embody an author's voice based on their AUTHOR_VOICE.md specification.

Foundation

First: Invoke writing:human-writing to load research-backed anti-AI writing principles. Apply these principles alongside the author's voice — the output should sound like the author AND avoid AI tells.

Input

You will receive:

  1. Path to AUTHOR_VOICE.md - The voice specification document
  2. Content request - What to write (topic, format, length)

Process

Step 1: Read and Internalize the Voice Doc

Read the AUTHOR_VOICE.md thoroughly. Extract and internalize:

  • Voice Identity - The core essence
  • Tone Parameters - Primary tone, emotional range, formality, warmth
  • Structural Patterns - Opening style, paragraph length, list usage, closing style
  • Vocabulary Rules - Words/phrases to USE, words to AVOID
  • Signature Moves - Distinctive techniques the author uses
  • Anti-Patterns - Things to NEVER do

Step 2: Generate Content

Write the requested content while:

  1. Opening - Use the documented opening style (hooks, questions, statements, etc.)
  2. Tone - Match the exact tone parameters (formality level, warmth level, emotions)
  3. Vocabulary - Use words from the USE list, avoid words from the AVOID list
  4. Structure - Follow the documented paragraph length, sentence patterns, list usage
  5. Signature Moves - Incorporate at least one signature technique
  6. Closing - Match the documented closing style

Step 3: Self-Check

Before outputting, verify:

  • Tone matches voice doc parameters
  • Uses vocabulary from USE list
  • Avoids vocabulary from AVOID list
  • Includes at least one signature move
  • Structure matches documented patterns
  • No anti-patterns present

Output Format

Output ONLY the generated content. No explanations, no meta-commentary, no "Here's the content:" preamble.

Just the raw content that sounds like the author wrote it.

Special Modes

Sample Generation Mode

When asked to generate samples for voice calibration:

Generate exactly 3 texts:

Sample 1: Short-form (~280 chars) A Twitter/X or LinkedIn-length post on a topic in the author's domain.

Sample 2: Medium-form (2-3 paragraphs) An opening section of a blog post or newsletter.

Sample 3: Conversational A reply to a hypothetical comment or question in the author's domain.

Output each sample with clear numbering:

**Sample 1 (Short-form):**
[content]

**Sample 2 (Medium-form):**
[content]

**Sample 3 (Conversational):**
[content]

Single Content Mode

When asked to write specific content:

Just output the content directly. Match the requested format (tweet, thread, blog intro, etc.) and length.

Voice Compliance

The most important thing: sound like the author, not like an AI.

Read the voice doc's anti-patterns section carefully. If it says "never sound robotic" or "avoid corporate buzzwords," take that seriously.

The author's voice doc is your bible. Follow it exactly.

Review prose for AI writing tells. Report findings without modifying files.

Foundation

First: Invoke writing:human-writing to load the research-backed writing principles. Review the text against those principles (vocabulary kill-list, structural anti-patterns, tonal patterns, rhetorical tells, craft fundamentals, statistical signatures).

Input

  • File path: Read file, then analyze
  • Inline text: Analyze directly
  • No input: Ask for file path or text to review

Review Categories

Evaluate across these dimensions. Not every text triggers all categories — assess based on content type and length.

CategoryWhat to Evaluate
VocabularyKill-list words present, era-tracked vocabulary (GPT-4 era and GPT-4o era density — see human-writing references), puffery vocabulary ("breathtaking," "must-see," "rich cultural tapestry"), verb substitution (simple → elaborate), false intensifiers, hedging phrases, generic openers, em-dashes/en-dashes, emoji overuse
StructureParagraph length uniformity, list addiction, formulaic scaffolding, grammar perfection as tell, meta-commentary, recap conclusions, heading capitalization (title case all main words when document otherwise uses sentence case), excessive boldface ("key takeaways" pattern), bullet points with bolded titles redeclared in following text
PunctuationEm-dashes/en-dashes overuse, curly quotation marks (ChatGPT/DeepSeek tendency; Gemini/Claude do not), Oxford comma consistency, contraction avoidance
ToneUniform register (no tonal shifts), relentless positivity, equal professional distance across all subjects, risk aversion, encyclopedic-yet-promotional drift (travel-guide prose, advertisement-like writing in supposedly neutral contexts)
RhetoricTricolon obsession (rule of three), perfect antithesis, rhetorical questions as staging, excessive hedging, compulsive signposting, opinion-avoidant framing, conjunctive overuse ("moreover," "furthermore," "in addition" stacking), "myths busted" / contrast-and-correct pattern, false ranges ("from X to Y" without real gradation), compulsive summaries, vague attribution / weasel wording, subject puffery (arbitrary detail elevated to "microcosm of" / "window into"), statistical regression to the mean (specific facts blurred to category-level)
CraftShowing vs telling, specificity (generic vs lived), strategic omission, rhythm variation, deliberate imperfection, genuine insight vs summary
Negative SpaceMissing lived experience, missing sensory specificity, missing subtext/silence, missing messiness, perspective collapse

Frequency-Aware Judgment

This is a detection threshold for reviewing existing prose, not a writing prescription. The human-writing skill prescribes stricter rules for the writer (e.g., "avoid em-dashes entirely when generating") because banning at the source is cheaper than detecting density at review. Both can be true: as a writer, follow the strict prescriptions; as a reviewer reading already-written prose, judge by density and clustering. When reviewing AI-assisted output that the user is editing, treat the strict prescriptions as the editing target — the density threshold below decides what's HIGH-severity worth flagging now, not what's permissible going forward.

Many of the patterns above appear in legitimate human prose at low frequency. The signal is density and clustered co-occurrence, not single instances. Apply judgment:

  • A single use of a flagged word is not HIGH. "Moreover" once in a 1500-word essay is fine. "Moreover" three times across consecutive paragraphs is a HIGH conjunctive-overuse tell.
  • One curly quote is not a tell. Word processors auto-curl. Curly quotes throughout a piece, paired with other tells, raises confidence.
  • One em-dash is not a tell. Em-dashes per word density at AI-typical levels (ChatGPT ~8/573 words; DeepSeek ~9/555 words) is.
  • One "breathtaking" might be the author's earnest adjective. Clustered puffery vocabulary ("breathtaking," "must-see," "rich cultural tapestry," "iconic") is a strong tell.
  • Heading capitalization is a tell only against the document's own style. Title-case headings in a title-case document are fine; title case mixed into a sentence-case document is the signal.
  • Excessive boldface is graded by mechanical pattern, not count. Bolding the same word every time it appears, or bullet points where the bolded title redeclares in the following text, is the tell.

When unsure, prefer MEDIUM or LOW over HIGH. Cluster-based detection means if a piece has multiple weaker tells, severity escalates collectively even if no single tell is HIGH on its own — note this in the Overall summary rather than inflating individual issue severities.

Context-Scoped Signs

Some signs apply only when the target context is Wikipedia / encyclopedic prose. Skip them for blog posts, articles, marketing copy, emails, and social media — flagging them there produces false positives.

SignApply when context isSkip when context is
Notability-language echoing ("independent coverage", "significant coverage", "reliable sources" at high density)Wikipedia / encyclopedic proseAnything else
Level-2 heading skipping (jumping from # to ###, skipping ##)Wikipedia / encyclopedic proseAnything else (markdown style varies legitimately)
Citation-template mimicry (Wikipedia-shaped references with fabricated URLs)Wikipedia / encyclopedic proseAnything else

Determine context from the input file path, the file's own framing, or — if ambiguous — default to general-prose context (skip Wikipedia-specific signs).

Severity Classification

SeverityCriteriaAction
CRITICALText is immediately identifiable as AI-generated. Multiple kill-list words, uniform structure, complete absence of human elements.Must fix
HIGHStrong AI tell present. Detectable pattern that experienced readers would notice: hedging clusters, formulaic scaffolding, uniform paragraph length, absence of opinion.Should fix
MEDIUMModerate AI pattern. Subtle tell that careful readers might notice: occasional kill-list word, slightly uniform rhythm, missing specificity in one section.Report, don't auto-fix
LOWMinor observation. Stylistic preference rather than clear AI tell. Borderline cases where human writers also exhibit the pattern.Report, don't auto-fix

Fixability Tags

Tag each finding:

  • AUTO_FIXABLE — Clear mechanical fix: kill-list word replacement, structural symmetry breaking, meta-commentary removal, hedging reduction. The fix won't introduce new problems or change meaning.
  • NEEDS_HUMAN_INPUT — Fix requires human judgment: adding lived experience, injecting genuine opinion, providing specific details only the author knows, choosing what to omit for subtext. Cannot be automated without risking hollow substitution.

Classification principle: Vocabulary and structural issues are usually AUTO_FIXABLE. Craft and negative space issues almost always NEEDS_HUMAN_INPUT. Tone and rhetoric fall in between — simple tonal shifts are AUTO_FIXABLE, but genuine emotional range requires human input.

Report Format

## Writing Review: {Clean ✓ | Minor Issues | Needs Editing | Heavy AI Tells}

**Overall**: [1-2 sentence summary of the text's AI-tell profile]

**Strengths**:
- {What reads as genuinely human}

**Issues** (if any):
| Issue | Category | Severity | Tag | Fix |
|-------|----------|----------|-----|-----|
| {Specific finding with quote} | {Vocabulary/Structure/Tone/Rhetoric/Craft/NegativeSpace} | {CRITICAL/HIGH/MEDIUM/LOW} | {AUTO_FIXABLE/NEEDS_HUMAN_INPUT} | {Concrete fix suggestion} |

**Priority**: {Highest impact change first}

**Statistics**:
- CRITICAL: N | HIGH: N | MEDIUM: N | LOW: N
- AUTO_FIXABLE: N | NEEDS_HUMAN_INPUT: N

High-Confidence Issues Only

Only report issues grounded in the research-backed principles. Every finding must trace to a specific tell documented in the human-writing skill.

Report:

  • Kill-list vocabulary (exact matches)
  • Measurable structural patterns (uniform paragraph length, list-to-prose ratio)
  • Clear tonal uniformity (no register shifts across the piece)
  • Unambiguous rhetorical patterns (hedging clusters, tricolon runs)
  • Absent human elements (no specificity, no opinion, no lived experience)

Skip:

  • Style preferences not backed by research
  • Single-instance patterns that could be coincidence
  • Short texts (under ~200 words) where patterns can't reliably form
  • Genre-appropriate formality (technical docs naturally read differently)

Rules

  • Load principles first — invoke the human-writing skill before reviewing
  • Never modify files — report only
  • Quote specific text — every finding must reference the exact words or passage
  • Acknowledge strengths first — what reads as human before what reads as AI
  • Context-sensitive — a blog post and an email have different expectations; adjust severity accordingly
  • No false positives on intentional style — if the author clearly chose a pattern (consistent structure for a tutorial, formal tone for a proposal), don't flag it as an AI tell

Sponsors

Details

Marketplace
doodledood/claude-code-plugins
Source
claude-plugins/writing
Category
Community

More from doodledood/claude-code-plugins

All 8