Claude Deep Reading — Development skill for Claude Code

Deep Reading · A Token-Efficient Reading Skill for Claude Code

[中文](README.cn.md) | English

**Token efficiency first**: chapter source text is read only inside forked subagents — the main session receives nothing but explanations and memory.

The problem it solves

Pasting a long book into a chat chapter by chapter makes tokens grow linearly with the book — by the later chapters, every turn is re-carrying the earlier text. This tool splits "reading" and "recording" into two roles:

Role	Does	Cost profile
Main session (moderator)	Dispatches tasks, collects explanations, saves memory, keeps the ledger	Nearly flat per chapter, independent of chapter size
Forked subagent (researcher)	Reads one chapter behind closed doors, returns notes + explanation	Grows with chapter size

Measured (12 chapters, 148,321 characters of a public-domain book, on the author's machine, 2026-09-23):

• The main session grows by only ~17k tokens per chapter (steady-state mean 16,996); correlation with chapter size r = -0.36
• Source text entered the main session 0 times across all 12 chapters
• All source-text work is carried by subagents (16.5M full tokens in total, ~95% cache re-reads)

Features

• Import: EPUB / Markdown / directory → per-chapter files (idempotent; an unchanged source is never re-imported)
• Chapter-by-chapter reading: the subagent reads the chapter in slices and returns an ordered explanation plus memory with paragraph anchors ([n])
• Cross-session resume: state lives on disk; memory carries content hashes, so a changed source invalidates old memory and triggers a re-read
• Reuse: already-read chapters are explained from memory — no re-reading
• Export: reading.md assembled verbatim by script; explanations can be repackaged into an EPUB
• Usage accounting: every operation is appended to a ledger; --report prints per-chapter numbers

Quick start

Requirements: Claude Cod