NovelToGame

Seven open-source agent skills for a coding agent you already use — Claude Code, Codex or Kimi Code — that turn a novel into a game. Give it a book and, if you have one in mind, a target platform or engine: it reads the whole text, writes a source bible in which every hard rule, character goal, turning point and signature detail carries the chapter or file location it came from, compares real directions and eliminates them with named hard vetoes, designs the world and the look, builds for the runtime the brief locks instead of falling back to a web page, and proves the result plays in one recorded run of six checks. Three modes: quick drafts the brief with defaults and stops only on blocking items, director stops at the concept for you to choose, and resume reads the progress file and continues from the last stage that genuinely finished. A run leaves five design documents, a runnable build and a QA record. There is no GPU, no hosted service and no bundled engine, and three adaptations ship as playable browser games with no install.

Who built itThe organization account that owns and maintains the repository, with five sibling projects listed in its README and a six-project overview on the organization site. All 138 commits carry one GitHub account, worldwonderer, while the author field alternates between pitechen on 54 of them and PiteChen on 81, and every commit comes from a single address. The contributor list holds one name. 55 commits carry a co-author trailer and every one of them names a Claude model — Opus 4.8 with a million-token context on 23, Opus 5 with a million-token context on 16, Fable 5.1 on 12, Opus 5 on 3 and Fable 5 on 1.

1. 01

   Sixty-five days, five releases, and one account on every commit

   The repository was created on 2026-07-18 and its oldest commit is titled Make novel-to-game portable across agent CLIs; the newest, on 2026-09-21, is chore: prepare v0.4.0 release (#65). Between them sit 138 commits — 72 in July, 45 in August, 21 in September — and five releases, every one of them stable: v0.1.0 on 2026-07-31, thirteen days after the repository appeared, then v0.2.0 on 2026-08-05, v0.3.0 on 2026-08-23, v0.3.1 on 2026-09-05 and v0.4.0 on 2026-09-21. The five tags match the five releases exactly, with no prerelease or draft among them, and the release titles carry the roadmap: evidence-first delivery and the third example at v0.2.0, replayable adaptation contracts at v0.3.0, a leaner skill package at v0.3.1, blind-tested skills and one run per example at v0.4.0. Around that sit 817 stars, 115 forks and a single watcher on 65 MB counted by GitHub as Markdown, with one open issue — an outsider offering to add the project to a plugin list. All 138 commits carry a linked account, all of them worldwonderer, and all 138 come from one email address; the author field alternates between pitechen and PiteChen. 55 commits carry a co-author trailer and every trailer names a Claude model.

2. 02

   What source-grounded means, line by line

   The claim is checkable because the artifact is a table. In the Journey to the West workspace, analysis/SOURCE_BIBLE.md (12,229 bytes) is a two-column list of facts and evidence, and the evidence column holds chapter numbers: the true fan quells fire with one wave, raises wind with two and brings rain with three, against Chapter 59; the Bull Demon King disguised as Bajie tricks the true fan back, against Chapter 61. The README calls this the citation layer: every hard rule, key character goal, turning point and ending, and signature anchor carries its chapter or file location, a work without chapters citing the file, and every row of the adaptation-boundary table citing its evidence. Facts are then labelled immutable, adaptable, open or conflicted, and anything the novel does not define is marked a design invention rather than smuggled in as fact. The concept stage answers from the other side, writing one line per candidate direction recording which of six named hard vetoes it triggered: a core loop that ignores the novel’s central tension; strip the proper nouns and a generic template remains; the player only spends resources to release a fixed plot. Project Plateau adds the condition under which its winner should be abandoned — falsified if position or timing cannot create a visibly better plate, or if recording never changes a later route or defense decision.

3. 03

   Six checks, one run, and the limitations written beside them

   The QA record is a small file with a strict shape. qa/verification.json binds six named checks — launch, render, input, coreLoop, outcome, restart — to a single complete run, and each one is only ever NOT_RUN, FAIL or PASS: unverified is not a pass, and an old PASS is overwritten by every rerun. The same file carries a completeRun block, an id with a clean-context flag, the terminal state reached, the state the restart landed on and a pointer to the evidence file, and a list of limitations where each entry is a scope with a reason rather than a verdict. The Jin Ping Mei record is the candid one: all six checks pass and the first limitation says the fast path picked the first feasible main action on every screen and reached one unstable ending, with other options and endings not exhausted. Each example keeps exactly one authoritative command for this. Project Plateau runs npm run verify, which drives the real build with keyboard and mouse events and writes the input trace and the verdict in the same run; Journey to the West runs python3 test/verify.py; Jin Ping Mei runs python3 test/verify_visual.py --write-evidence. A pull request on 2026-09-20 deleted the second checker that had grown beside each of those three, on the rule that one check belongs at the real boundary.

4. 04

   Three checked-in workspaces, three locked runtimes

   The examples are whole adaptation workspaces rather than screenshots. Each holds PRODUCT_BRIEF.md, analysis/SOURCE_BIBLE.md, concepts/CONCEPT.md, design/GAME_DESIGN.md, design/ART_DIRECTION.md, build/BUILD_BRIEF.md, build/app/ and qa/verification.json, and the novels ship with them: 2.1 MB of Journey to the West, 2.1 MB of Jin Ping Mei beside a 9,858-byte Python script that expurgates it, and 451,824 bytes of The Lost World, each with its edition, coverage and rights status in source/SOURCE.md. What differs is the runtime the brief locked. The two Chinese examples are dependency-free static apps, and the Jin Ping Mei build carries the largest files in the repository — 1.1 MB of data.js and 706 KB of engine.js — while a 9,792-byte build/art/generated-art.json records all 41 of its images as generated with the image tool built into Codex, and the original cover and five portraits are pinned by SHA-256 hashes. Project Plateau, the 3D one, is a Three.js and Vite application of 160 files where the art is generated rather than drawn: seven .mjs scripts of 15 to 34 KB build the .glb terrain and vegetation libraries and a 14,804-byte asset-ledger.json accounts for them. The paperwork moves the other way, 92,095 bytes of game design in Jin Ping Mei against 10,787 in Project Plateau.

5. 05

   Trimming the skills, and measuring the trim blind

   The skills went through the same evidence discipline they impose on their users. Three pull requests in the last four days before v0.4.0 cut dead weight out of the package — rules each skill stated twice and hedges repeated several times inside one file — for the reason the repository wrote down when trimming its validator: a skill package pays for every line twice, once in the repository and once in the agent’s attention budget. The prose trim was measured rather than argued: a blind A/B on an invented novel fixture, three generations per arm, judged blind by three models on five axes. The first pass took the package from 1,089 lines to 1,060 and the trimmed arm won 18 of 27 pairwise comparisons on game-concept (6.87 against 7.13) and 16 of 27 on game-art-direction (7.00 against 7.27). The second pass is the interesting one, because it lost where it was tried next: game-world-design scored 6.84 as the control against 6.38 slim, six pairwise wins out of 27, a regression signal, and was reverted, while the neutral result in art direction — 7.07 against 7.24, 14 of 27 — was kept. The same window cut the line budget from 1,900 to 1,300, so the roughly 1,120 lines in the package became the ceiling; a fortnight earlier the validator had gone from 968 lines to 760 the same way, by folding duplicated schema checks into shared helpers and deleting dead code and a whole scan.

6. 06

   The reversals the repository kept on the record

   Four failures are written down as plainly as the successes. The art direction was redone: an early pass produced dark, antique and photoreal experiments that were rejected and replaced by one luminous 2D language across all 41 runtime images. The cover then went wrong a second time in a way worth recording — the bright title image was never lost, but a dark group-portrait title screen added later stayed in force after the portrait itself was judged to have made the game worse and withdrawn, and both the README card and the QA screenshot captured that state; the overlay was removed and the screenshot re-recorded in the same run. The deploy pipeline had a fault that only appeared in production: the Vercel projects had come to define their own root directory while the workflow also changed into that directory first, so the command line resolved the path twice and failed before upload; the fix runs it from the repository root, and a regression test rejects per-job working directories. And Project Plateau’s own product brief records that the first measured run crossed the whole route in 55.2 seconds, falsifying the planned five-to-eight-minute session, so the product boundary was cut to one to three minutes rather than padding the route with waits. The one number the README refuses to give is its own cost: the progress files record neither wall-clock time nor token spend.