Book Research

Book Research

Supply evidence rather than collecting links. Separate facts, proxies, inferences, and hypotheses, and write every material finding into book.db with provenance so it survives the conversation.

This skill inherits every rule in AGENTS.md. It never overrides human creative authority or database safety. Route through $book-guide first for any mixed request.

Research modes

Pick the mode from the request; do not load the heavy framework for a small question.

1. Market whitespace — find defensible, evidence-backed opportunities for a new book and series. This is the heavy mode; load references/market-whitespace.md only here.
2. Comparable titles and reader analysis — build a comparable-title dataset; analyze demand, satisfaction, supply, reader jobs, and tension pairs.
3. Factual story research — subject expertise, historical/technical accuracy, setting and profession detail.
4. Claim verification — check a specific in-world or marketing claim against sources.
5. Authenticity and sensitivity — representation, lived experience, and harmful-similarity risk.
6. Publishing and route-to-market — traditional, independent, serial-first, direct, hybrid; format and channel fit.

For modes 2–6, apply the discipline below without reading the full whitespace framework.

Progressive loading

references/market-whitespace.md is the full deep-research prompt (demand × dissatisfaction × supply gap × author advantage × packaging × renewable engine × route to readers). Read it only for substantial market-opportunity work. A quick fact-check or single comparable must not load it.

Non-negotiable research rules

1. Date-stamp the work and state market/geographic scope.
2. Cite every material claim (title, site, publication/observation date, access date, link or stable id).
3. Never convert rank, ratings, review counts, views, followers, library holds, or search interest into unit sales. These are distinct signals. (Mirrors AGENTS.md write rule 9.)
4. Triangulate: a major conclusion needs ≥3 independent signals, ideally different source classes.
5. Use multiple horizons: recent velocity (12 mo), current structure (36 mo), durable demand (5–10 yr), incoming competition (18–24 mo ahead).
6. Sample failures and abandoned series, not only breakouts.
7. Respect paywalls, copyright, privacy, robots rules, and platform terms. Paraphrase reader observations; keep provenance.
8. Do not imitate a hit's surface. Extract the underlying reader promise and engine.
9. Return confidence levels and risks. Do not promise success.

Persistence into book.db

When the work is substantive, persist it (see AGENTS.md “Research market white space”):

• one research node for the study; store the full report as its current research text version;
• one source node per material source, with date, territory, source class, and limitations in data;
• one concept node per opportunity finalist;
• statement rows linking sources to concepts with predicates supports, contradicts, qualifies, signals_demand, signals_risk (mode='evidence', real confidence);
• entry rows for assumptions, missing data, experiments, and decisions.

Concepts stay idea or candidate. Never mark a concept canon or approved without an explicit human choice (AGENTS.md write rule 7). Rank is not sales; store each signal as what it is.

If book.db is unavailable, also return a machine-readable object (sources, opportunities, evidence_links, assumptions, experiments) for later import.

Definition of done

A research turn is complete when the author received evidence-backed findings with clear confidence, every material claim is cited, sales and proxies were kept separate, concepts remained candidates, and durable conclusions were saved into book.db with provenance — or a machine-readable fallback was returned.