For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 391 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
A piece by Thariq at Anthropic, widely read in May 2026, lands an observation that is easy to miss when the default has been markdown for years: format choice is a signal to the reader about how the output is meant to be consumed. Markdown signals sequential reading. HTML signals scanning, interaction, direct edit. SVG signals structural at-a-glance comprehension. Choosing one is choosing a reading mode for whoever is on the other end.
The McLuhan riff is intentional. The shape of the artifact is part of the message; ignoring the shape and letting it default to whichever format the writer happened to be trained on leaves real leverage on the table.
Once the framing is named, the missing variable becomes obvious: who reads this output, and how do they need to engage it?
The default-markdown-for-everything failure mode is the failure to ask the consumer-question. Once asked, the format follows.
In his framing, the writer of the artifact and the surface that delivers it are the same agent making one choice: write markdown, or write HTML, or write SVG. The choice happens at write-time and is baked into the artifact.
There is a generalization. The writer can keep to one durable shape, and the surface can render to whatever the consumer needs. Two layers, not one:
Most knowledge systems collapse these into one. The artifact is born in markdown and stays in markdown until forever, regardless of consumer. Or the artifact is born in HTML and is never readable outside a browser. The writer-layer and surface-layer separation refuses both collapses. It costs more (a renderer per surface) and gains the freedom to reshape per consumer without rewriting.
There is a consumer Thariq's piece does not engage at depth: another agent. When the consumer is itself a model, sequential markdown is often correct again, because models read sequentially within a context window. But "often correct" is not "always correct". A model that needs to compare ten options would benefit from a structured shape (JSON, a table, a graph) for the same reason a human comparing ten options benefits from side-by-side HTML. A model that needs to navigate a large corpus would benefit from indexed shape, not linear shape.
The agent-consumer case is where the writer/surface separation pays the largest dividend. The same durable artifact, rendered into a structured shape for the agent and a narrative shape for the human, lets one source serve both consumers without either getting the other's affordances. Whether to build that rendering layer for any given system depends on whether two consumer-questions actually fire on the same artifact. Increasingly, in workflows where humans and agents touch the same content, they do.