Every post on this site has a unique image generated from its title. Here's why - and how.

If you have noticed that the header images on this site do not look like the usual product-blog image set, that is not accidental.

No stock photos, no AI-generated illustrations, and no generic gradient cards with the title dropped over a royalty-free texture.

Each one is a deterministic geometric composition: a unique arrangement of shapes calculated directly from the title of the post.

Same title, same image, always.

The concept: deterministic Suprematism

The algorithm is rooted in Suprematism, the early 20th century Russian art movement Kazimir Malevich built around a simple proposition: pure geometric form, freed from representational meaning, could carry its own visual logic.

Malevich’s compositions were not random. They had weight, tension, axis, and balance. The shapes related to each other according to internal rules. Nothing was arbitrary.

What I built is a computational version of that idea. The algorithm takes a blog post title, converts it into a numeric seed through a hash function, then uses that seed to generate a Suprematist composition. The angle of the master axis, the position and colour of each element, the tension between the shapes: all of it is a mathematical output of the words in the title.

The image is the post’s signature.

Why this approach

The obvious alternatives all have problems.

Stock photography is generic by definition: the same images appear everywhere, communicate almost nothing specific about the content, and signal that you reached for the easiest option.

AI-generated illustration is more interesting, but it introduces randomness. Without a careful system, the images across a hundred posts start to feel incoherent. They also look like AI-generated images, which now carries its own aesthetic tax.

Generic design cards, title over gradient with maybe an icon, are fine for utility. They have no character. They do not create the kind of visual identity that helps a reader recognise your work in a feed before they have read the name.

The deterministic approach solves all three problems.

Each image is unique. Each image is consistent within the system while still varying in composition. The entire library, across however many posts, shares an aesthetic DNA because the same rules are applied to different inputs.

The connection to SR-SI

There is a connection here to Simulated Recall via Shallow Indexing, the AI context methodology I published in February.

SR-SI is also deterministic. The same context, indexed correctly, produces consistent and coherent AI outputs, not random ones that depend on how you phrased the prompt that day. The architecture ensures the same input can reliably reconstruct the right output conditions.

The image generator applies the same logic to a different problem. The title is the input, the composition is the output, the rules are fixed, and the result is always derivable and reproducible.

Deterministic systems are more honest than random ones. They commit to a relationship between input and output, then hold to it.

What it looks like in practice

The algorithm works in layers. A master angle, the compositional axis, is seeded first, and everything else orients to it. Then comes a centre of gravity, a primary anchor shape, medium blocks, needle-thin slicers, and finally satellites: small floating accents distributed along the axis vector.

The palette is fixed: the four colours from this site’s identity system. The shapes are rectangles, squares, and circles; the rules for how they relate to each other are fixed, and the title determines which specific combination of those possibilities gets activated.

Type “The Handoff Gap” into the generator and you get one specific image. Always the same one. Change it to “The Coordination Tax” and you get a different composition: different angle, different weight distribution, different colour balance.

Both are Suprematist, both are from the same system, and neither looks like the other.

The export problem, automated away

The export problem was the last piece. A nice image system does not help much if every post still needs a manual export pass before it can appear properly on the site, in social previews, and across other surfaces.

Right now the system is fully automated. Every post on this site has its own generated image ready to be served without a manual design decision from my end. I loved designing the earlier hero images, but realistically, if I wanted to keep up the cadence, something had to give. The choice was either dropping cover images altogether or finding a way to automate them while preserving some character.

I chose the latter because I could not reliably generate images on demand without burning through too many tokens.

The end result is the thing I was trying to build: a content identity system that scales without degrading.