process-next-book

name	process-next-book
description	Process and adversarially fact-check the next ready non-fiction book in this repository from claim through research, structured content, reputation-blind rating, narration, local audio, validation, and queue completion. Use only when the user explicitly asks to run or process the book queue.

Process Next Book

Complete exactly one queued book. Treat AGENTS.md, config/rating.json, and config/audio.json as the live instructions; do not copy or replace their rules here.

Workflow

1. Confirm the current directory contains AGENTS.md, bookflow, and data/queue.json. Stop if any are missing: this skill must operate in the 5MinBooks repository. Read AGENTS.md completely, inspect the working tree, and preserve unrelated or concurrent changes.
2. Run ./bookflow next --claim. Process only the book it returns. If none is ready, follow the discovery procedure in AGENTS.md, initialise the strongest well-supported candidate, and claim it.
3. Run ./bookflow check <book-id> before editing. Use its output and ./bookflow queue to avoid redoing complete, current work.
4. Research online and complete the book and author records to the source and coverage standards in AGENTS.md. Use broad, claim-appropriate sources, keep coverage truthful, and correct existing work rather than merely listing its defects.
5. Complete content.json, tags, catalogue metadata, and evidenced graph relationships. Keep fields distinct, concise, critical, and useful for the reader's decision.
6. Read config/rating.json. Assess every configured dimension — the ideas as delivered on the page and the reading craft of their delivery — with a source-backed rationale. Score craft from documented reading-experience and reception evidence, never the rater's taste. Exclude author prestige, popularity, awards, aggregate review scores, and personal agreement. Record coverage uncertainty as confidence; do not alter component scores because access was partial. Run ./bookflow rate <book-id> --write to calculate the total. Never choose or hand-edit the total.
7. Read docs/review-method.md and run its fact-check and adversarial passes separately. Verify identity and edition metadata, fidelity to the book, consequential external claims, counterevidence, citation entailment, rating discipline, product fit, and simple language. Correct defects in the records and scripts. Set the dated quality review to passed only when every check is true; never use review notes to excuse a known error.
8. Derive the configured narration scripts from content.json, compare adjacent lengths with the loss test, and set the shortest sufficient recommended level. Complete every configured duration. Keep research sources, coverage, methodology, workflow and production notes out of the narration; start with the book itself and end on content or a reading decision. Use ./bookflow audio <book-id> for local audio; do not reproduce its word-count, ordering, freshness, or TTS logic manually. Read docs/tts.md, listen to the shortest output, and check the title, author, technical terms, abbreviations and numbers. Add only verified corrections to the shared pronunciation dictionary, then regenerate affected audio.
9. Run ./bookflow check <book-id>, then ./bookflow check and ./bookflow build. Fix every in-scope failure. Mark the queue entry done only when the repository definition of done is met. If genuinely blocked, record the reason and set the entry to blocked instead of inventing data. When the owner requested commits, commit this completed book before claiming the next one; never include an unfinished next book in that commit.

Handoff

Report the book, score and confidence, recommended listening level, research coverage, source count, audio generated or skipped, validation results, and any unresolved limitation. Do not commit or push unless the user asks.