very-long-text-summarization

name	very-long-text-summarization
license	Apache-2.0
description	Summarizes very long texts (books, handbooks, biographies, codebases) using hierarchical multi-pass extraction with cheap model armies. Produces structured knowledge maps, not just summaries. Use when processing 50+ page documents, professional handbooks, career biographies, or any text too large for a single context window. Activate on "summarize book", "summarize handbook", "long document", "extract knowledge", "distill text", "professional biography". NOT for short text summarization (<10 pages), real-time chat summarization, or code documentation (use technical-writer).
allowed-tools	Read,Write,Edit,Bash,Grep,Glob
argument-hint	[file-path-or-url] [output: summary|knowledge-map|skill-draft]
metadata	{"category":"Content & Writing","tags":["very","long","text","summarize-book","summarize-handbook"],"pairs-with":[{"skill":"research-craft","reason":"Long-text summarization supports research synthesis of academic papers and books"},{"skill":"llm-router","reason":"Multi-pass summarization uses model routing to assign cheap models for extraction passes"},{"skill":"cost-optimizer","reason":"Hierarchical summarization of large texts requires cost-aware model selection"}]}
category	AI & Machine Learning
tags	["summarization","long-context","text-processing","llm","compression"]

Very Long Text Summarization

Processes texts too large for a single context window using hierarchical multi-pass extraction with armies of cheap models. Produces structured knowledge maps, indexed summaries, and skill drafts — not just prose compression.

Decision Points

1. Output Mode Selection

If user wants quick understanding for meeting prep:
  → Use SUMMARY mode (2-pass: Haiku army → Sonnet synthesis)
  
If user needs machine-readable knowledge for downstream processing:
  → Use KNOWLEDGE-MAP mode (2-pass with full structured extraction)
  
If user wants to convert handbook expertise into Claude skill:
  → Use SKILL-DRAFT mode (3-pass: add Opus refinement)

2. Document Size Assessment

If < 20 pages (< 40K tokens):
  → Skip this skill, use direct summarization
  
If 20-200 pages (40K-400K tokens):
  → Standard 3-pass pipeline, 4K chunks with 500-token overlap
  
If 200+ pages (400K+ tokens):
  → Large document mode: 8K chunks, 1K overlap, parallel batch processing

3. Quality vs Cost Trade-off

If budget < $0.05:
  → Haiku-only mode: single pass extraction, no synthesis
  
If budget $0.05-0.20:
  → Standard mode: Haiku extraction → Sonnet synthesis
  
If budget > $0.20 AND output is skill-draft:
  → Premium mode: Add Opus refinement pass

4. Iteration Threshold

If knowledge map has < 80% concept coverage on first pass:
  → Run second extraction pass with focused prompts on gaps
  
If extracted processes have < 3 decision points each:
  → Re-extract with decision-focused template
  
If < 5 failure modes identified in technical text:
  → Run failure-mining pass with anti-pattern detection

Failure Modes

Schema Bloat

• Symptoms: Knowledge map has 50+ concepts, 20+ processes, extraction takes 10+ minutes
• Diagnosis: Over-extraction without prioritization, treating every detail as key knowledge
• Fix: Add importance scoring in Pass 1, filter to top 20% concepts in Pass 2

Chunking Amnesia

• Symptoms: Duplicate concepts with different names, missing cross-references, fragmented processes
• Diagnosis: No overlap between chunks, semantic boundaries ignored
• Fix: Increase overlap to 25% of chunk size, use semantic chunking on headings

Attention Dilution

• Symptoms: Generic summaries, missed domain expertise, no decision trees extracted