Use these while you build

Reading a reference once and then hunting for it six weeks later is the normal way this material gets wasted. Every note here is downloadable as markdown with stable headings, and each file carries its own shader, specifically so it can live next to your code and be retrieved instead of remembered.

The problem with knowing something once

Say you worked out that a neon core has to be multiplied past 1.0 rather than blended toward white. Two months later you are in a different project, at a different hour, and you reach for mix(hue, white, t) again, because that is what the instinct says. The knowledge existed. It just was not anywhere your hands were.

Same story with the numbers. Nobody remembers that a 24-step soft shadow costs about ten points of a WebGL2 frame, or that a lit forehead sits 0.13 from a cream wall in RGB. Those are exactly the facts that decide an implementation and exactly the ones that evaporate.

Put them where the agent looks

If you build with a coding agent, the useful move is to keep this material in the repository the agent already reads, and index it so a question retrieves the passage rather than the agent guessing from training data. The difference in practice: 

# without it
"how should I do the glow?"
  -> a plausible one-layer exp() falloff, no numbers, no reason

# with it retrievable
"how should I do the glow?"
  -> three falloffs at 4r / 1r / 0.3r, gains 0.4 / 2 / 6,
     and the reason the core must exceed 1.0 rather than lerp to white

That is not a smarter model. It is the same model with the answer in front of it.

What that looks like concretely

  1. Keep the notes as markdown in the repo, not in a wiki the agent cannot open.
  2. Give each one a stable title and stable section headings, so a link to a claim keeps working.
  3. Index the folder for search. Semantic plus keyword beats either alone.
  4. Write the retrieval step into your agent's standing instructions, so it queries before it answers rather than after it is wrong.
  5. When you learn something new, add it in the same pass. A note written a week later is a note that never gets written.

The method, if you want the whole thing

This site is one shelf out of a much larger network built the same way: research, decisions, and the reasons behind them, committed where the work happens so a fresh session can pick up cold. I wrote up how that is put together, including the parts that failed, at Build With the Memory.

Worth saying plainly, and it is on the front page too rather than only here: these notes were produced by agents and then adversarially verified, and the verification is the part that matters. Several claims on this site were wrong on the first pass and got caught by a checking step, including one where the demo looked convincing and proved nothing. The method is the interesting artifact, not the prose.

Getting the notes

Two ways, and both are files rather than pages:

Each file carries the note's prose, its rules of thumb, and the complete shader that runs in its demo, self-contained. They are generated from the pages themselves, so what you download and what you read are the same thing.