Decide. Render. Judge. Revise.

One call walks eleven stages. The model is used where judgment lives; everything that touches pixels is deterministic code. That split is why a post can be revised without being re-rolled.

1. 01BriefmodelThe model reads the business context and distills what this post is for: goal, audience, the one message, and the proof that backs it.
2. 02GatescodeDeterministic gates kill vague briefs before any pixels exist. No concrete message, no supported claim, no post.
3. 03AnglesmodelThree to five candidate points of view. Angles that would work for any competitor fail the swap test and are dropped.
4. 04CopymodelHook options of three to seven words, a platform-native caption, and a call to action in the reader's voice.
5. 05FormatcodeContent picks the format: a number wants a stat card, a testimonial wants a quote card, a real moment wants a photo hook.
6. 06ImagerymodelReal brand photos lead. When the pool is spent, the image model paints a name-free, text-free atmospheric scene instead.
7. 07The dealcodeA dealer deals one of five layout archetypes and a typography voice, deterministically per brand and post. No roulette, no repeats.
8. 08RendercodeA real layout engine composites the text over the scene. Pixels come from code, so warped text and garbled logos are structurally impossible.
9. 09JudgemodelA vision judge checks four hard gates and eleven weighted dimensions, then hands down a verdict with its reasons on the record.
10. 10RevisemodelBelow the bar, only the weakest stage re-runs while everything else holds. Bounded, surgical, and never worse than it started.
11. 11MasterscodeThe settled spec renders true 4:5 and 9:16 masters, separate renders, not crops. Caption, hashtags, score, and the spec ride along.

Every visual decision is a serializable spec: the archetype, the frame geometry, the type system, the palette, the words. The renderer never improvises. Same spec, byte-identical PNG, every time. Below: a real spec and the master it rendered, judged SHIP at 95.

spec.json · abridged

json

{
  "format": "photo-hook",
  "copy": { "onImageText":
    "Roasted this week, not last spring" },
  "layoutSpec": {
    "archetype": "split-block",
    "frame": { "kind": "split", "axis": "y",
               "ratio": 0.55 },
    "anchor": "center-start",
    "lead": "image"
  },
  "typeSystem": {
    "display": { "family": "Oswald", "source": "brand" },
    "body":    { "family": "Lato",   "source": "brand" }
  },
  "colors": { "primary": "#2B1D16",
              "accent":  "#C96F2E" }
}

// judge: { "verdict": "SHIP", "score": 95 }

Because the spec is the design, an edit is a diff. A revision can fix the headline while the photo, frame, and palette hold verbatim.

Every render faces a vision judge before it returns. Four hard gates come first; failing any one is a no-ship regardless of the numeric score.

Design quality is graded on its own ladder, TEMPLATE to COMPETENT to STRONG to GREAT, so a post with fine words and a formulaic layout still gets sent back. The revise loop is bounded: at most two passes, each re-running only the weakest stage, with a no-regression floor so a fix never ships something worse than it started with. The one exception is deliberate: when a user dictates exact words, their words win, and the judge's dissent is recorded instead of obeyed.

A feed dies when every post is the same shape. Layout is dealt from five archetypes, type from six voices, and the deal is deterministic per brand and post: the same post re-renders identically, the next one refuses to repeat it, and two businesses in the same town stop wearing the same face.

Brands with real fonts keep them; the voices dress everyone else. Real dealt renders are on the homepage gallery and the image world.

A carousel is planned as an arc: a hook that earns the swipe, a rehook that stands alone, body slides that each carry one idea, and a closing ask. The dealer picks one frame family for the whole deck, and variation comes from each slide's role, not from per-slide randomness. Randomness reads as a tell; the judge punishes it. The whole eight-slide deck is on the carousel world.

• claude-opus-4-8decides the brief, angles, and copy; judges every render
• claude-sonnet-4-6runs the fast text loop: scoring, rewriting, planning
• gemini-3.1-flash-imagepaints name-free, text-free atmospheric scenes
• satori + resvgcomposite the design deterministically, in the Worker
• Cloudflare Workersone worker owns the API, MCP, renderer, and this site

The same engine also cuts short-form video: narrative beats, motion b-roll, captions timed to a real voiceover. It runs Taka's reels in production today and joins the public API when it clears the same bar the rest of this page describes. Watch one on the video world.