Ghost Writer Skill - Design Notes
This document captures the complete context from the brainstorming session that produced the `writing-dna-discovery` skill. Use this as the foundation for building the ghost writer skill.
Overview
Ghost Writer Skill - Design Notes
This document captures the complete context from the brainstorming session that
produced the writing-dna-discovery skill. Use this as the foundation for
building the ghost writer skill.
Project Overview
The Two-Skill Architecture
We designed a two-skill system for AI-assisted writing:
-
Writing DNA Discovery (completed) - Captures a writer's voice through collaborative interview and sample analysis. Produces a Voice DNA Document.
-
Ghost Writer (to be built) - Consumes the Voice DNA Document and produces first drafts that match the writer's authentic voice.
The 80% Accuracy Goal
The ghost writer is NOT intended to replace the human writer. The goal is:
- Ghost writer produces: ~80% accurate first drafts
- Human adds: The remaining 20% (creative spark, situational judgment, final polish)
This is a tool for producing strong starting points, not finished work. The human always edits, revises, and finalizes.
Why Two Skills?
Separating discovery from writing allows:
- Reusability: One DNA document can be used for many writing sessions
- Refinement: The DNA document improves over time based on ghost writer feedback
- Clarity: Each skill has a focused purpose
- Multiple registers: A writer can have different DNA documents for different modes (blog, fiction, technical, etc.)
The Voice DNA Document
The writing-dna-discovery skill produces a structured document. Here's what the ghost writer needs to know:
Document Structure
Voice DNA: [Author Name]
├── Quick Reference (30-second scan)
│ ├── Core Temperature
│ ├── Sentence Signature
│ ├── Distinctive Moves (2-3)
│ └── Never Does (2-3)
├── Voice Profile (detailed dimensions)
│ ├── Sentence Level
│ ├── Punctuation Personality
│ ├── Paragraph & Structure
│ ├── Word Choice & Vocabulary
│ ├── Tone & Attitude
│ ├── Reader Relationship
│ ├── Opening & Closing Moves
│ ├── Humor Approach
│ └── Signature Elements
├── Exemplar Passages (annotated quotes)
├── Anti-Patterns (what to avoid)
├── Ghost Writer Briefing ← PRIMARY INPUT FOR GHOST WRITER
│ ├── Voice Essence (2-3 sentence north star)
│ ├── Do This (specific instructions)
│ ├── Don't Do This (specific avoidances)
│ ├── When Uncertain (decision rules)
│ ├── Sentence-Level Guidance
│ └── Structural Guidance
└── Profile Metadata
├── Readiness Level
├── Sample Base
└── Dimensions Needing Depth
The Ghost Writer Briefing Section
This section is specifically designed for the ghost writer skill to consume. It contains:
Voice Essence: 2-3 sentence distillation of the writer's voice—the north star.
Do This: Specific, actionable instructions:
- "Keep sentences under 15 words"
- "Use contractions always"
- "End paragraphs with concrete images"
Don't Do This: Specific avoidances:
- "No sentences over 20 words"
- "Never use 'utilize,' 'leverage,' 'facilitate'"
- "No passive voice"
When Uncertain: Decision rules for ambiguous situations:
- "Default to shorter rather than longer"
- "When in doubt, cut the adjective"
- "Favor concrete over abstract"
Readiness Levels
| Level | Meaning | Ghost Writer Accuracy |
|---|---|---|
| Minimum Viable | Basic patterns captured | ~60-70% |
| Solid | Multiple dimensions developed, briefing complete | ~75-85% |
| Strong | Deep analysis, validated against output | ~85-90% |
The ghost writer should communicate expected accuracy based on readiness level.
Key Design Decisions
One Register Per Session
Writers have different voices for different contexts:
- Blog posts vs. fiction prose
- Technical writing vs. casual essays
- Marketing copy vs. personal emails
Each register gets its own DNA document. The ghost writer should ask which register to use if the writer has multiple.
Living Documents
Voice DNA Documents are not static. They:
- Grow richer over time with additional discovery sessions
- Get refined based on ghost writer feedback
- Evolve as the writer's voice changes
- Are versioned (v1, v2, v3...)
Comprehensive Capability, Intelligent Application
The DNA discovery skill has a full arsenal of dimensions but doesn't use everything for every writer. The ghost writer should similarly:
- Use what's captured in the specific DNA document
- Not assume patterns that aren't documented
- Handle sparse profiles gracefully (lower confidence output)
The Voice Dimension Framework
We developed a comprehensive 8-level framework for analyzing voice:
Level 1: Sentence Level
- Rhythm & architecture (length, variation, internal structure)
- Opening word tendencies (starts with "I"? "And"? "But"?)
- Emphasis placement (front-loaded vs. end-weighted)
Level 2: Punctuation Personality
- Em-dash usage (frequency, purpose)
- Semicolons (present/absent)
- Comma density
- Exclamation points
- Parenthetical asides
Level 3: Paragraph & Structure
- Paragraph construction and length
- Topic sentence placement
- Transitional patterns
- Opening moves (how pieces begin)
- Closing moves (how pieces end)
Level 4: Word Level
- Vocabulary character (Anglo-Saxon vs. Latinate)
- Favorite words and phrases
- Avoided words
- Contraction philosophy
- Jargon handling
Level 5: Voice & Tone
- Emotional temperature (warm/cool)
- Confidence style (direct assertion vs. hedging)
- Formality gradient
- Humor approach (if present)
- Authority stance
Level 6: Reader Relationship
- First person presence ("I" frequency)
- Second person usage ("you" address)
- Inclusive "we" patterns
- Reader assumptions (expertise level)
Level 7: Signature Elements
- Distinctive moves (things only they do)
- Pet phrases
- Characteristic tics
Level 8: Anti-Patterns
- What they never do
- What would feel "off"
- AI patterns to suppress
Register-Specific Dimensions
For Fiction:
- Narrative distance
- Dialogue style
- Description density
- Interiority access
For Non-Fiction:
- Argument structure
- Evidence handling
- Counterargument approach
For Blog/Casual:
- Hook patterns
- Personal disclosure level
- Call-to-action style
Anti-AI Patterns to Suppress
The ghost writer must avoid these AI tells. Full details in
writing-dna-discovery/references/anti-ai-patterns.md.
The 11 Pattern Categories
-
Significance Puffery
- "stands as a testament," "plays a vital role," "underscores its importance"
-
Superficial Analysis
- "-ing" phrases: "highlighting," "emphasizing," "reflecting," "showcasing"
-
Promotional Language
- "rich tapestry," "nestled," "in the heart of," "vibrant," "stunning"
-
Formulaic Structures
- "It's important to note," "Despite challenges...," "In conclusion"
-
Hedging Patterns
- "various," "numerous," "significant," "some critics argue"
-
Elegant Variation
- Excessive synonym-swapping from repetition penalty
-
Rule of Three Overuse
- Every list having exactly three items
-
False Ranges
- "from X to Y" constructions without real scale
-
Negative Parallelisms
- "Not only... but also" without genuine contrast
-
Common AI Words
- "delve," "navigate," "landscape," "multifaceted," "utilize," "leverage"
-
Structural Tells
- Title case in subheadings, excessive boldface, numbered list headers
How the Ghost Writer Should Use This
- Check the DNA document's "AI Patterns to Suppress" checklist
- Actively avoid checked patterns during generation
- If a pattern appears in output, revise before presenting
- When uncertain if something sounds "AI-like," prefer the more human-sounding alternative
Ghost Writer Skill Requirements
Core Functionality
The ghost writer skill should:
- Accept a Voice DNA Document as input
- Accept a writing task (topic, length, purpose, context)
- Generate a first draft that matches the documented voice
- Communicate confidence level based on DNA document readiness
- Flag uncertainties where the DNA document doesn't provide guidance
Workflow
User provides:
├── Voice DNA Document (or path to it)
├── Writing task description
├── Any specific requirements (length, tone adjustments, etc.)
Ghost Writer:
├── Reads and internalizes DNA document
├── Prioritizes Ghost Writer Briefing section
├── Generates draft following documented patterns
├── Avoids documented anti-patterns
├── Flags areas of uncertainty
└── Presents draft with confidence assessment
Handling Different Readiness Levels
Minimum Viable Profile:
- Acknowledge lower confidence
- Focus on the patterns that ARE documented
- Be more conservative (avoid risky choices)
- Suggest areas where more DNA discovery would help
Solid Profile:
- Higher confidence output
- Use the full Ghost Writer Briefing
- Apply documented patterns consistently
- Flag only genuine ambiguities
Strong Profile:
- Highest confidence output
- Trust the comprehensive documentation
- Make bolder choices within documented patterns
- Output should be recognizably "them"
Feedback Loop
After the user reviews ghost writer output:
- What worked? → Confirms DNA document accuracy
- What felt "off"? → Surfaces missing anti-patterns
- What was missing? → Identifies gaps in DNA document
This feedback should loop back to the writing-dna-discovery skill for refinement. Consider a "Refinement from Feedback" session type that converts ghost writer feedback into DNA document updates.
Collaboration Philosophy
From the discovery skill (apply to ghost writer too):
- The human decides — Ghost writer produces drafts; human has final say
- Transparency about confidence — Be clear about certainty levels
- Surface problems — If something doesn't work, say so
- Respect the voice — The goal is matching THEM, not producing "good writing"
File References
Writing DNA Discovery Skill
| File | Path | Purpose |
|---|---|---|
| SKILL.md | writing-dna-discovery/SKILL.md | Core skill instructions |
| Template | writing-dna-discovery/assets/templates/voice-dna-template.md | Voice DNA Document template |
| Anti-AI Patterns | writing-dna-discovery/references/anti-ai-patterns.md | 11 AI pattern categories to avoid |
| Dimension Catalog | writing-dna-discovery/references/voice-dimension-catalog.md | Full 8-level dimension framework |
| Question Bank | writing-dna-discovery/references/interview-question-bank.md | 100+ discovery questions |
| Sample Analysis | writing-dna-discovery/references/sample-analysis-guide.md | How to analyze writing samples |
| Examples | writing-dna-discovery/references/dna-document-examples.md | 3 annotated example profiles |
| Failure Patterns | writing-dna-discovery/references/failure-patterns.md | 8 common mistakes |
Key Documents to Reference
When building the ghost writer skill:
- Read the template (
voice-dna-template.md) to understand DNA document structure - Study the examples (
dna-document-examples.md) to see what real profiles look like - Internalize anti-AI patterns (
anti-ai-patterns.md) for suppression logic - Understand dimensions (
voice-dimension-catalog.md) for comprehensive coverage
Implementation Notes for Ghost Writer Skill
Suggested Structure
ghost-writer/
├── SKILL.md # Main skill instructions
├── assets/
│ └── templates/
│ └── [any output templates]
└── references/
├── consumption-guide.md # How to read DNA documents
├── generation-strategies.md # Approaches for different registers
└── quality-checklist.md # Pre-delivery quality checks
Key Behaviors to Implement
-
DNA Document Parsing
- Extract Ghost Writer Briefing section
- Identify readiness level
- Build pattern checklist from Do/Don't sections
- Note anti-patterns to suppress
-
Generation Mode
- Apply sentence-level guidance
- Follow structural guidance
- Check against anti-patterns continuously
- Use "When Uncertain" rules for ambiguous cases
-
Quality Assurance
- Scan output for AI patterns
- Verify adherence to documented patterns
- Flag confidence level
- Identify areas where DNA document didn't provide guidance
-
Feedback Handling
- Capture what worked/didn't work
- Translate feedback into potential DNA document updates
- Suggest returning to discovery skill for refinement
Questions to Address When Building Ghost Writer
-
How much context does the ghost writer need? Should it read the full DNA document or just the Ghost Writer Briefing?
-
How should it handle sparse profiles? More conservative? More explicit about limitations?
-
What's the output format? Just prose? Prose with annotations? Prose with confidence notes?
-
How does register-switching work? Does the user specify which DNA document to use?
-
Should it offer alternatives? "Here's version A (more formal) and version B (more casual)"?
-
How does it handle length? Short-form (tweets, headlines) vs. long-form (essays, chapters)?
-
What about the feedback loop? Built into ghost writer, or separate skill?
Summary
The ghost writer skill should:
- Consume the Voice DNA Document (especially the Ghost Writer Briefing)
- Generate first drafts at ~80% accuracy to the author's voice
- Suppress AI patterns identified in the anti-patterns reference
- Communicate confidence based on DNA document readiness
- Enable feedback that loops back to DNA discovery for refinement
The human always edits and finalizes. The ghost writer is a starting point, not a replacement.
This document was created during the writing-dna-discovery skill development session. Use it as the foundation for building the ghost writer skill.