extract-memory

Analyze a collection of documents and extract a personal writing style profile into a Claude Code memory file.

What This Skill Does

1. Reads all documents from a given directory or file
2. Analyzes writing patterns, style, tone, vocabulary, themes
3. Generates a structured writing style profile
4. Saves the profile to a memory file (default: ~/.claude/writing-style.md)

Supported Input Formats

• .md — Markdown files
• .txt — Plain text files
• .pdf — PDF documents (read via the Read tool)
• .lakebook — Yuque exports (auto-detected, will run /convert-lakebook first)

Workflow

Step 1: Identify Input

Determine what the user provided:

• Directory path: Read all supported files in the directory
• Single file path: Read that one file
• .lakebook file: First convert to Markdown using python3 ~/Desktop/lakebook-converter/convert.py, then process the output directory

Step 2: Sample Documents

For large collections (20+ files):

• Read all files but sample strategically — pick diverse documents by length and topic
• Aim to read at least 15-20 substantial documents (>500 chars each)
• Prioritize longer, more expressive documents over short notes

For small collections (<20 files):

• Read all files completely

Step 3: Analyze Writing Style

Examine the documents and extract the following dimensions:

Structure Patterns

• How do they open articles? (Hook style, question, statement, story)
• What's the typical article structure? (Sections, numbering, flow)
• How do they close? (Summary, call-to-action, question, open-ended)

Language Characteristics

• Vocabulary level and preferences
• Use of metaphors and analogies (what domains do they draw from?)
• Sentence length patterns (short and punchy? Long and flowing?)
• Use of rhetorical devices (repetition, contrast, parallelism)
• Multilingual patterns (code-switching, foreign terms)

Tone & Voice

• Formality level
• Emotional range (analytical, passionate, humorous, serious)
• Relationship with reader (teacher, peer, friend, mentor)
• Confidence level and how they handle uncertainty

Thinking Patterns

• How do they approach problems? (First principles, analogy, data-driven)
• Common frameworks and mental models they use
• How do they handle disagreement or controversial topics?

Content Themes

• Recurring topics and interests
• Core beliefs and values that surface across multiple pieces
• What they reference frequently (books, people, concepts)

Formatting Habits

• Use of bold, italic, headers
• List styles (numbered, bullet, emoji)
• Use of tables, code blocks, quotes
• Paragraph length preferences

Step 4: Generate Profile

Create a structured Markdown file with the analysis results. Use this template:

# Writing Style DNA — [Name/Alias]

## Structure Patterns
[findings]

## Language Characteristics
[findings]

## Tone & Voice
[findings]

## Thinking Patterns
[findings]

## Content Themes
[findings]

## Formatting Habits
[findings]

## Core Identity
[one paragraph summary of the writer's essence]

## One-Line Summary
> [single sentence that captures their writing DNA]

Step 5: Save & Report

1. Write the profile to the target memory file
2. Report to the user:
   • How many documents were analyzed
   • Key findings summary (3-5 bullet points)
   • Where the file was saved
   • How to use it: "You can now tell Claude to read ~/.claude/writing-style.md before writing tasks"

Usage Examples

/extract-memory ~/Desktop/lakebook-converter/output/
/extract-memory ~/Documents/blog-posts/
/extract-memory ~/Desktop/export.lakebook
/extract-memory ~/Desktop/my-article.md

With Custom Output Path

If the user specifies where to save:

/extract-memory ~/Documents/posts/ --save-to ~/.claude/my-style.md

Default save location: ~/.claude/writing-style.md

Important Notes

• Do NOT store the full document content in the memory file — only the distilled style analysis
• Keep the output under 3KB to avoid excessive token consumption when loaded as context
• If a memory file already exists at the target path, ask the user if they want to overwrite or merge
• The analysis should be specific and actionable, not generic — include actual examples from the user's writing