The "making of": AI as editor & the mnemonic narrative

[GregStanton]

Hi all!

This repository's WebGL2 & GLSL primer was written during an era of LLMs, which raises the question: Was this written with AI? I'll describe the writing process here, as well as the unique format that (currently) appears to require human expertise.

AI as editor

I used AI as a book editor, rather than a co-author. I wrote the background section, the cards (sometimes with AI brainstorming), and the solution code for the projects. I have experimented with using AI to draft this type of content, but the guide's unique form seems to require a human expert, at least for now.

Since the projects are established rites of passage for graphics programmers, Gemini did save me time by drafting project specifications. It also deserves credit for the "programmable geometry pipeline" line, which emphasizes the primer's procedural focus. As an editor would do for a book jacket, Gemini also wrote initial drafts of the repo description for the sidebar, and the "Community & next steps" section.

The mnemonic narrative

I refer to the form of the primer as a mnemonic narrative, as it balances the continuous nature of storytelling with the discrete nature of spaced-retrieval practice. It involves a logical plot with a background (prerequisites), characters (concepts), and a climax (applications). While the cards are presented through an educational narrative, turning this into a mnemonic medium requires each card to be carefully crafted, so the cards can be reviewed out of order (according to retrieval difficulty), and even interleaved with other subjects, as per standard spaced-retrieval workflows (e.g., with Anki).

Why human expertise is required (seemingly, for now)

I did experiment with using AI to generate cards, including coaching it on prompt-writing principles, but teaching AI to do this effectively has been fairly hard. While it can be helpful to ask AI for a seed-stage draft to work from, I think I always or nearly always ended up rewriting or restructuring any cards it created. This was true even for Gemini Guided Learning, which is explicitly customized for teaching.

I suspect that part of the issue is that writing good card prompts is an advanced skill on its own. Writing them in the context of a narrative adds a significant layer of difficulty on top. It requires a constant awareness of the full context in which each card is introduced, which needs to be balanced with learning principles like atomicity, elaboration, and context independence. Training data for this educational format may also be scarce, as it extends beyond typical fact-based card decks.

Feedback on the primer's format and its execution

The primer is open source, and the cards are written in HTML, which Anki understands. This means that after you learn a card, you can throw it into Anki for regular practice. If anyone has feedback on this format, or its execution, please share it here!

Thanks to everyone,
Greg