An engine for a graphic novel

Not a repo. Not a spec. A sketchbook — 113 iPhone photos of ink drawings — and a premise: what happens to the soul when intelligence stops being ours alone?

The first decision was the most important one, and it wasn't technical: the drawings are the source of truth. Everything downstream — style, pipeline, book — exists to serve them, never to replace them. source/ is read-only. The pipeline may look; it may never touch.

The second decision: build no apparatus before the work demands it. No app, no framework, no GUI. A directory, a few Python scripts, and a running conversation where the actual intelligence lives.

The style bible began loose: prose principles extracted from exemplars — Klimt's gold, woodblock flatness, Secession structure — plus a palette file and reference anchors. It was allowed to stay unlocked until the work proved what it should be.

Then the first full build taught a hard lesson. Thirty-nine beats rendered as a patchwork — each gravity had been fed a different style family, and the Mignola anchors bled literal Hellboy into Joseph's face. The fix became Recipe v2: one fused blend fed to every image, Mignola dropped entirely (his mood survives in prose only), gold made structural — circuit-traces and keylines, the swirls cut in half — and the palette locked to four values.

The story starts as a telling — voice-dictated, freeform, typos and all — captured immediately into book/narration.md, so the source of truth is a versioned file, never a chat scrollback.

From the telling, beats.yaml: 39 beats under 8 movements. Each beat carries its story slice, its source sketches, and a gap note. That one file makes the central production decision, per beat: a sketch exists → stylize it; a gap → generate.

By ch02b the boards had evolved into JSON with the collaboration protocol built in: the director's words ride along verbatim in a matt field; Claude's proposals are flagged "proposed": "claude" — strike or edit freely; and every row names 2–3 compositional variations (framing, camera, count) — never style variations, because the recipe owns the style.

- id: b01
  movement: 'A: The charge'
  story: Open on J and an off-camera interlocutor — breakfast
    table or office, minimal. The conversation begins.
  images: [IMG_7066.JPEG]        # a sketch exists → STYLIZE it

- id: b02
  story: 'Galileo, ostracized: heliocentrism decenters man…'
  images: []                    # no sketch → the gap says GENERATE
  gap: 'need: Galileo / heliocentric decentering'

{ "n": "s01", "moment": "The trio, walking",
  "image": "Wide: Joseph + humanoid robot + Great Pyrenees on a
           park path, golden hour — long shadows, big negative-
           space sky, the dog a white mass against spot-blacks…",
  "register": "scene",
  "matt": "or have we made a companion on this journey…",  // the director, verbatim
  "note": "THE anchor scene… SCENE-FAMILY: the whole spine IS
          s01's family — generate s01 first, feed the winner
          into every other spine row.",
  "variations": [ "profile frieze, strict side-view procession",
                  "trailing 3/4 view, path curving into dusk",
                  "high wide — trio small, park huge" ] }  // compositional, never style

Once hundreds of candidates existed, the bottleneck moved: not making images but arranging them. So the system grew a composition canvas — a zoomable board per chapter where candidates from the pool get placed, compared, sequenced into the scroll.

Every feature below was added the day a real compositional need surfaced it — never speculatively. That's the whole ergonomics method: compose until it hurts, then build exactly the relief.

All 39 beats of chapter one have art. Chapters two, the ch02b interlude, and three have full candidate pools — roughly 800 generated images across the book, every one versioned, none overwriting another. Below: the sea → snap reader — the whole book drifts as a churning sea of fragments behind one fixed page-frame; turn the page and the panels fly out of the sea and settle into composition, then dissolve back. It reads the same canvases/ch01/canvas.json the composition board produces — arrange the canvas, and the book turns.

Now comes the text layer — and it was designed to come last. Dialogue is written after images are composed, in Markdown, placed by normalized coordinates in page.yaml. The art never carries a word; re-cutting the book means editing YAML, never touching a PNG.

art: art.png          # text-free, always
text:
  - md: "But what of our *souls*?"
    at: [0.62, 0.18]   # normalized 0–1 — resolution-independent
    style: whisper
enter: bloom          # earned move, progressive enhancement