Book Review

• Length: short (1,500–3,000 words), medium (3,000–6,000), deep (10,000–20,000). Default: deep.
• Voices: The user can optionally specify voices:
  • 3 voices — skip voice selection, go straight to drafts.
  • 2 voices — accept them, auto-select a third to maximize diversity, present all three for approval.
  • 1 voice — single-voice mode: skip voice selection (Stage 3) and parallel drafts (Stage 4). Write one draft directly as 05_synthesis.md (replacing the normal multi-voice synthesis of Stage 5), then proceed to fidelity check and polish. Do not write 03_voice_selection.md or any 04_draft_*.md files.
  • 4+ voices — ask the user to pick 3 or fewer.
  • 0 voices (default) — auto-select 3 voices.
• Resume: If the user says "resume from [stage]", read prior saved outputs and pick up from that stage.

Stage 3: Voice Selection

Read 01_comprehension.md and 02_critical_analysis.md.

Voice Roster (suggestions, not exhaustive):

Any author works — the roster is a starting point, not a constraint.

Selection criteria:

• Maximize diversity — pick from different categories.
• Match to material — empirical claims benefit from analytical voices; human experience from literary ones.
• Productive tension — at least one voice somewhat adversarial to the book's style or assumptions.
• Leverage strengths — pick voices whose characteristic way of noticing surfaces things others miss.

Reasoning must be specific: Not "ACX because analytical" but "ACX because this book makes three falsifiable claims in chapters 4, 7, and 11 that deserve stress-testing."

Write 03_voice_selection.md:

# Voice Selection: [Book Title]

## Voice 1: [Name]
**Category:** [category]
**Reasoning:** [specific reasoning for this book]

## Voice 2: [Name]
**Category:** [category]
**Reasoning:** [reasoning]

## Voice 3: [Name]
**Category:** [category]
**Reasoning:** [reasoning]

CHECKPOINT: Present the picks to the user and wait for approval. They can swap voices, accept, or request re-selection. Do not proceed until approved.

Stage 4: Voice Drafts

Dispatch 3 subagents in parallel via the Agent tool, one per voice. Each subagent's prompt must include:

• The voice to write in
• The paths to 01_comprehension.md and 02_critical_analysis.md
• The path to the book file
• The length target
• The output file path

If parallel execution is unavailable, run drafts sequentially.

Output: 04_draft_<voice_name>.md per voice. Normalize voice names for filenames: lowercase, replace non-alphanumeric characters with underscores, collapse consecutive underscores.

Length per draft: deep 5,000–8,000 words, medium 2,000–4,000, short 1,000–2,000.

Instructions for each draft:

• Comprehension and critical analysis are raw material — don't cover everything. Follow what's most interesting through this voice's lens.
• Structure is free-form. Follow the thinking, not the chapter order.
• The voice shapes what you emphasize and notice, not just how you write. Different voices ask different questions.
• Engage with ideas directly — argue, extend, complicate. Don't just describe.
• Use the critical analysis as a starting point but go beyond it. Do something with the weaknesses.
• Don't hedge. Have actual opinions. Take positions.
• Voice is a way of thinking, not a costume. Don't reduce it to surface tics.
• Different voices should notice different things. Don't force them to cover the same ground.

CHECKPOINT: After all drafts complete, notify the user. They can optionally read the drafts and give feedback before synthesis.

Stage 5: Synthesis

Read all voice draft files plus 01_comprehension.md and 02_critical_analysis.md.

• Identify which voice's way of engaging produced the most insight — not just prettiest prose.
• Adopt that as the primary voice.
• Extract from all drafts: strongest arguments, sharpest observations, best examples, interesting tangents, unique insights.
• Write a new, coherent review that weaves the best material from all three. This is a genuine rewrite, not cut-and-paste.
• The final piece should feel like one unusually perceptive reviewer. No seams.
• Free-form structure. Let the piece find its shape.

Length: deep 10,000–20,000 words, medium 3,000–6,000, short 1,500–3,000. Short means picking 2–3 most important threads, not compressing everything.

Write 05_synthesis.md.

Stage 6: Fidelity Check

Read 05_synthesis.md and the book file.

• Extract every factual claim the review makes about the book.
• Verify each against the source text. For long books, work claim by claim — search for relevant passages rather than holding the entire book in context.
• Categories of error: misrepresentation, fabrication, misattribution, false absence, exaggeration.
• Fix errors in place. If a correction changes the force of a critical argument, adjust the surrounding argument.
• Preserve voice and flow. Corrections should be invisible.

Write 06_fidelity_checked.md (corrected review) and 06_fidelity_log.md:

# Fidelity Check Log: [Book Title]

**Total claims checked:** [number]
**Issues found:** [number]

## Issue 1: [brief description]
**Category:** [category]
**Claim:** [what the review claimed]
**Explanation:** [what was wrong]
**Correction:** [what was changed]

This is the most important quality gate. A stylistically brilliant review that misrepresents the book is worse than useless.

Stage 7: Polish

Read 06_fidelity_checked.md.

• Light pass. The review is already substantively complete.
• Catch AI-isms: generic phrases ("it's important to note," "delve into," "at the end of the day"), excessive hedging, passive voice overuse.
• Sharpen bland vocabulary.
• Do NOT alter voice, structure, or humor — those are baked in.
• Do NOT add introductions, conclusions, or framing that weren't in the synthesis.

Write 07_final.md.

CHECKPOINT: Print the final review to the conversation and notify the user it's saved to 07_final.md.

Error Handling

• If the book file cannot be read (unsupported format, too large, missing), tell the user and stop.
• If a stage produces incomplete output, retry once. If it fails again, show the user and ask how to proceed.
• If a subagent in Stage 4 fails, the other drafts are still usable. Offer to re-run the failed draft or proceed with what succeeded.
• File-based state provides implicit resumability. If any stage fails, completed outputs are saved. The user can say "resume from [stage]."

Common Mistakes

• Voices as costumes — "Use footnotes like DFW" is wrong. The voice should change what you notice and emphasize, not just surface stylistic tics.
• All voices covering the same ground — The whole point is that different voices surface different things. If all three drafts discuss the same chapters in the same order, the voices aren't working.
• Synthesis as cut-and-paste — The synthesis must be a genuine rewrite, not paragraphs stitched together from three drafts.
• Skipping fidelity check — Never skip Stage 6. It's the most important quality gate.