Teaching LLMs New Tricks

Azoth — open source library.

• Winter 2023–24: built it.
• When AI hit: shelved it.
• Nov 2025: AI changed the math, and brought back Azoth!

Creating a dashboard for a real estate brokerage — Google Looker — couldn't find the right thing in the "vending machine."

Visually intensive components — just write HTML and CSS?

Combining a brokerage with a property developer.

Added integration analysis — high user touch.

JSX everywhere; the AI defaults to React — same syntax, wrong semantics.

Azoth — without vDOM, controlling renderer, no JS creating the DOM — <p>Hello</p> just returns a p tag!

First instinct: treat the new framework like a new framework (take an adjacent concept); just use it a bit.

This backfired! Use the language, then it confuses the LLM.

"Don't Say…, Say Instead" Map

Don't confuse our innovation with the incumbent's.

Don't onboard the LLM: hire it as the new senior maintainer.

AND… they're curious about your tech!

Shared mental model: a system metaphor that is shared; start a conversation.

Don't describe in negative space against something else — talk about the thing as it is.

Azoth JSX is real DOM

• What does it compile to? Doesn't interfere with JavaScript.
• Takes the HTML and makes a template (the JS creates no DOM).
• Doesn't use a factory function to make the DOM.
• Add how to bind, small helper functions — deduplicated.

Like the opposite of React — what you opt into with React you opt out of with Azoth.

Subtraction, not replacement

Don't replace the vDOM, subtract it.

The leap: closures and DOM — reduce to a DOM problem.

component = constructor, literally.

LLM, hey, THIS IS DOM! Reach for a tool that makes sense for the problem.

Azoth basically accepts everything that JSX already has. No proprietary primitives, just JS.

Notice the shape — this was the unlock. REMOVE the abstraction.

New mapping is SUBTRACT and unlock, not "don't say, say this instead."

Prompts as Navigation

Prompts aren't engineering instructions. They're navigation.

Push away from what is being subtracted, and pull towards what's unlocked.

Avoid NEGATION, like "Don't think about a pink elephant."

Law of AI Era Innovation

• Context — what you carry into the conversation.
• Corpus — what's already in the model — your budget! Be smart about this!

Subtract your abstractions, unlock the corpus.