What source analysis is responsible for
NovelToGame drafts PRODUCT_BRIEF.md first, then runs novel-game-analyze to produce SOURCE_BIBLE.md, and only then compares gameplay, designs and builds (see the full novel-to-game workflow). This step builds the factual layer: which rules exist, who wants what, where things happen and how the world responds. It does not judge the prose, choose a genre, or write state keys, scene_id or interface structure; those belong to GAME_DESIGN. Later stages may not silently rewrite either upstream file.
If a brief exists, read it first: its rating, core fantasy and any locked genre decide which actions, relationships and set pieces get evidence first.
The *Journey to the West* case on the zenstory.ai homepage is a useful reference: the rule “the westward road is constrained” rests on Flaming Mountain blocking the only road west (chapter 59), and the adaptation goal is to reopen the road with no somersault-cloud skip (that goal belongs to the later design step). Below, an original novel shows the same evidence work.
Fix the source range and coverage first
If you already have a writing project or a full deconstruction library, reuse it and return to the text only for key evidence. Run the whole-book pipeline only when all you have is the raw novel: set boundaries → extract unit by unit → pass the coverage gate → aggregate in layers → check back against the text → compress and deliver. An analysis covering only the first three chapters or a quick preview can help you judge the opening, nothing more.
The source range records the edition, adaptation range, unit boundaries, missing parts and rights boundaries. Acquisition details such as download dates and checksums stay out unless reproduction needs them.
Track coverage per unit. A chapter read with no relevant finding still counts as covered; unread or failed units are marked separately. The table header states the number of source units and how many succeeded, and the covered set must match the source units. Non-overlapping chapter batches may run in parallel, but they share one extraction contract, aggregation and fact-checking stay sequential, and nothing falls back to sampling. Unless you explicitly limit the adaptation to one volume or excerpt, an analysis that does not cover the whole text does not move on to concept design.
What to record per unit: only facts that affect play
The protocol lists ten review dimensions: objective change, action candidates, rule evidence, actors, space, items and abilities, knowledge gaps, pacing and feeling, audiovisual anchors, and evidence and boundaries. They are questions to ask while reading each unit, not a form to fill in for every chapter. Record only evidenced findings that affect player actions, world responses or audiovisual identity; leave no placeholder fields and do not retell the plot.
Three of them decide whether there is anything to play later:
- Action candidates: actions a player could repeat with choice, feedback or room for skill, including source-backed orders, commissions and errands;
- Rule evidence: triggers, costs, exceptions, counters and observable consequences;
- Actors: goals, default behaviour, resources, limits, and what they can order or take on.
For space, record ways in and out, thresholds, landmarks, risks and what can change; for knowledge gaps, who knows, misbelieves, hides or lies, and through what carrier. Hard rules, key character goals, grounds for the player role, turning points and endings, and signature anchors need chapter locations, and inferences must be labelled. When the source language differs from the design language, quote the original, add the translations a decision needs, keep one shared glossary, and do not swap rites or religion for another culture’s near-equivalent label.
Original example: chapter 1 of Night Letters from Elm City
*Night Letters from Elm City* is a novel invented for this article; its characters, city and rules are original. Line numbers mark evidence locations.
L1 Elm City has nine gates. After the drum sounds the third watch, every gate is barred and no one passes.
L2 Only someone holding a red-lacquered wooden token can have the north gate opened. The token must be returned to the post station before the hour of the Dragon next morning, countersigned by the gatekeeper. If it comes back late, the station issues no tokens for a month.
L3 Su He has run letters for three years. The route she knows best goes from the post station through the alley behind the dye works to the north gate, half the distance of the main street.
L4 Near the end of the second watch, Du Qi, who keeps the apothecary, presses a wax-sealed letter on her: “Someone at the farm outside the walls has a high fever. The prescription is inside. It must arrive tonight.”
L5 Station master Liang Yan hands her a red token. “I’ve sent word to Old Wen at the north gate to keep it open for you.”
L6 Su He reaches the north gate just as the third-watch drum dies away; the bar is already down. Old Wen says no one from the station has come tonight.
L7 Old Wen checks the token, then the wax seal. “I go by tokens, not by word of mouth. If you want the gate open, write in the gate ledger what you’re carrying.”
L8 Su He does not know what the letter says. She hesitates, then writes “prescription”.
L9 Night rain has soaked the field paths outside the walls. She sinks into the mud twice, and by the time she is back at the north gate the sky is pale and the token is spattered with mud.
L10 Old Wen countersigns the ledger but keeps the token. “I’ll take it to the station myself before the hour of the Dragon.”
Below is an evidence table written by hand for demonstration only. It is not the skill’s native output or the result of a run:
| Dimension | Record and location |
|---|---|
| Rule evidence | All nine gates barred after the third watch (L1); a red token opens the north gate (L2, L7); return and countersign before the hour of the Dragon, or no tokens for a month (L2), not seen enforced here |
| Action candidates | Running letters against the curfew, for three years (L3, L6, L9); showing a token (L7); declaring what is carried (L7, L8); returning the token (L2, L10) |
| Actors | Su He: knows the routes, not the letter’s contents (L3, L8); Old Wen: trusts tokens, not words, volunteers to return the token (L7, L10); Liang Yan: says he sent word (L5) |
| Space | Post station → alley behind the dye works → north gate (bar, ledger) → muddy field paths (L3, L6, L9) |
| Knowledge gaps | “Prescription” is only Du Qi’s claim (L4), entered in the ledger unverified (L8); the two men disagree on whether the station sent anyone (L5, L6) |
| Audiovisual anchors | The third-watch drum (L1, L6), the mud-spattered token (L9), the wax seal (L4) |
From evidence to gameable rules and player verbs
Do not map narrative elements straight onto genres. For recurring actions, rules, relationships, places, information gaps and emotional promises, ask four things: What decision can the player make? How does the world respond? What state carries into the next step? Does repetition produce learning or different results? A one-off climax stays a signature set piece, not the core loop.
“Running letters against the curfew” rests on three years of deliveries, and route, drum and mud all affect whether she arrives in time, so it is a system candidate. “Declaring in the ledger what you carry” is Old Wen’s condition for opening the gate (L7) and makes someone leave a checkable record without knowing the truth. It stays as a candidate too, though only the whole book can show whether the condition always applies. “The station’s message never arrived” happens once. Only when “an order is not the same as a job done” holds repeatedly in the source can it become a system candidate; one undelivered message cannot justify a delegation system. Record it as something to confirm across the book, and ask who gave the order, who carried it out, which people and places it passed through, what evidence it left and when it becomes a problem again.
Adaptation boundaries must be a table using only four labels: immutable, adaptable, open (open design space) and conflicted (conflicting evidence). These are this example’s judgments within the excerpt:
| Item | Boundary type | Evidence |
|---|---|---|
| Nine gates barred after the third watch; a red token opens the north gate | immutable | L1, L2, L7 |
| Token returned before the hour of the Dragon, or no tokens for a month | immutable | L2 |
| The back alley is half the distance of the main street | adaptable | L3 |
| What the letter really says | open | L4, L8 |
| Whether the station sent anyone to the north gate | conflicted | L5, L6 |
| Whether Old Wen returns the token on time | open | L10 |
Both statements about the message are character claims, not world facts. Keep both pieces of evidence and do not decide for the novel who lied. Nor may attributes migrate between neighbouring characters or objects. While the player role is still open, compare which actions and decisions Su He, an original character, the post station as a faction, or an observer could reach.
Write down what you still do not know
The excerpt cannot answer the following. These go into design gaps or a to-check list instead of being filled in on the fly:
- What the letter says, and why Du Qi insists it arrive tonight (L4 and L8 offer only claims);
- Whether the month-long ban is ever enforced, and whether Old Wen returns the token on time (L2 only states the rule; the excerpt ends at L10);
- Whether the other eight gates have exceptions (L1 says only that all are barred);
- Whether the lost message recurs across the book.
When aggregating across the book, cite where a conclusion is first established, how it typically operates, its key counterexamples and its final payoff, so an opening-only pattern does not become a hard rule; check numbers, ability limits, relationships, deaths and endings against the text. With only this excerpt, mark the bible as a partial analysis and list unread chapters as not reviewed. Unless you explicitly limit the adaptation to this chapter, it does not move on to concept design (for choosing a range, see how much of a novel a game should adapt).
After the concept stage selects a passage, the skill runs again to append a fact | evidence location | boundary type table for it to the bible, including the clothing, weapons, build and position of every on-screen character, with “not specified in the source” where the text is silent. The excerpt never says what Su He wears, so you cannot give her a straw raincoat and call it the original. Once the bible is complete, choose gameplay from the source’s repeated actions.
FAQ
Can I adapt from a deconstruction that only covers the first three chapters?
It only helps you judge the opening. For a whole-book adaptation, coverage must match the source units, so read the missing chapters first, or state explicitly that only those chapters are being adapted and record that in the bible’s source range.
Can systems the novel never describes, such as numeric progression, go into the bible?
Not as source facts. When a locked genre needs a system the novel does not provide in systematic form, note under design gaps that the source has no systematic equivalent and the system is a downstream invention. The actual rules are left to the later design stage.
How is this different from a novelist’s story bible?
A writer’s story bible keeps stable facts, current state and knowledge boundaries consistent while drafting. A SOURCE_BIBLE serves adaptation decisions: it records what a player could do, how the world responds and what must not change, with source locations for key facts. It holds no downstream game design and does not fill in the novel’s setting.
Do it with the skill
Use $novel-game-analyze to turn the full novel I provide (or my existing full deconstruction library or writing project) into analysis/SOURCE_BIBLE.md, reading PRODUCT_BRIEF.md first if it exists. Start with the source range and per-unit coverage, then record world rules, playable verbs, spaces, character will, information gaps and visual anchors with source locations. Write adaptation boundaries as an “item | boundary type | evidence location” table using only the four boundary labels. Keep inferences, character claims and direct source statements apart, and put anything the source leaves undefined into design gaps. Do not choose a genre or write state keys or interface structure at this stage; if the source is not fully covered, report the gap instead of moving on to concept design.
More of the method: Where source analysis sits in the novel-to-game pipeline