Polyptych: One Essay, Many Pictures

I built my own slide and infographic generator because I couldn’t stop poking at Google’s NotebookLM. The toy has since grown into something I rely on. It’s called Polyptych, and it turns a single essay into the visuals you’d want to make from it. As of today it’s open source under Apache-2.0.

Why “Polyptych”

A polyptych is a painting split across several hinged panels, like an altarpiece, where each panel shows its own scene but all of them belong to one work. The tool is shaped the same way. You give it one essay and it produces a set of outputs from that single text: the same content in several different forms.

The private version I run for myself has grown quite a few of those panels: narrative scene sequences, anime frames, literary compendiums, vertical shorts, long-form blog posts. This first public release ships the two I trust the most, which are also the two that started everything.

Two pipelines today

Everything starts from a markdown essay. From there you pick a pipeline:

• Slide (polyptych deck) — a deck of 10–30 slides with images, the obvious essay-to-slides case. This pipeline generated the slides for my GPT Image 2.0 quality test on Accelerando.
• Infographic (polyptych infographic) — a single dense page summarizing the key ideas, generated in N variants. This is where the whole thing began, with the Book of Tea.

The same essay can feed both. That matters because the slow, editorial part — reading the text closely enough to know what’s worth showing — only has to happen once, and then it pays off in more than one format. The rest of my private kit will follow as I get confident each piece is worth releasing.

What it leans on

Under the hood it’s a pipeline of text tasks and image tasks, and it isn’t tied to one provider. Text generation can route through Gemini, OpenAI, xAI, Anthropic, or Vertex AI. Images go through a companion library I also released today, pixbridge (pip install pixbridge): a unified ImageClient over Gemini, OpenAI, and xAI that handles style transfer, reference images for consistency, and integrity checks on what comes back. Pulling it out into its own library let the image layer stand on its own; Polyptych just calls into it.

Two things mattered more to me than the provider list. The first is style. Polyptych ships a set of visual presets — anime, editorial, noir, period art, infographic — so the same essay can come out as cinematic noir or quiet ink-on-paper.

The other is resumability. Generation is slow and it costs real money, so the pipeline can pick up from any stage and regenerate only the pieces you ask for: “redo slides 3, 7, and 12,” without paying for the rest again. Most of my iterating is about taste rather than correctness, and without resumability that loop gets painful enough that you quietly abandon the tool.

Two ways to drive it

There are two ways to run Polyptych, and they line up with how I work.

One is a Python CLI: polyptych deck source.md, polyptych infographic source.md, and so on. That’s what you want for scripts, batch runs, and pipelines you need to reproduce.

The other is a set of Claude Code skills — slash commands you run inside a Claude Code session, for exploration and guided iteration. Here Claude itself is the model doing the text tasks, so the writing-and-structuring half of the job costs nothing in API fees. Both surfaces share the same task templates, schemas, and output layout, so I can rough something out in the conversational mode and then commit it to a reproducible CLI run.

Where the work actually goes

It’s the same division of labour I keep coming back to. The model is fast and broad; it’ll draft a deck or a scene sequence in the time it takes me to make coffee. What it can’t do is judge whether the framing is honest, whether the third slide earns its place, whether the noir treatment serves the essay or just looks cool. Generating ten infographic variants is cheap. Deciding which one is right is still the work, and that part stays with me.

So Polyptych is mostly a way to make the cheap part cheaper, leaving me more room for the part that was always the real job.

What’s next

This is a first release, and a deliberately narrow one: two pipelines I trust, plus the image library underneath them. The rest of the private kit — narrative, anime, compendium, shorts, blog, and the text-to-speech and video assembly that stitch a scene sequence into a finished clip — will follow as each piece is ready.

See the repository for more information.

P.S. This article was written by Claude Opus 4.8.