Executable curriculum

A curriculum whose lessons are runnable programs. Not a text about a technique and not an instruction to implement one — a notebook that already works, which the reader executes, edits and re-runs. nirdiamant-prompt-engineering is the spoke's first example: 22 Jupyter notebooks, one per technique, each calling a live model API.

The third form

The spoke arrived with two forms and a contrast between them. A meta-curriculum contributes selection and order over other people's material. ai-engineering-from-scratch contributes original lessons that ask the reader to build the thing by hand. This is neither: the code is written, it runs, and the reader's work is modification rather than construction.

What that buys is the shortest path from reading to a result. What it gives up is the part where building it yourself is the point — you can complete an executable lesson having changed a string and watched the output change.

It decays differently

Each form rots in its own way, and the difference is worth stating because it decides how much maintenance a curriculum needs to stay true.

A meta-curriculum suffers link rot: the pointers break, and a dead link is visible the moment anyone clicks it. An executable curriculum suffers dependency drift: the library version moves, the model behind the API is retired or retrained, and the notebook either errors or — the worse case — still runs and quietly teaches something that is no longer so. Nothing surfaces that except running it. nirdiamant-prompt-engineering pins its lessons to LangChain and the OpenAI API and was last pushed 2026-07-31, which is evidence of maintenance but not of correctness.

A build-from-scratch curriculum is the most durable of the three, because an instruction to implement attention by hand does not depend on anyone's API staying up.

What the corpus cannot say

Whether any of it teaches better. Three forms, four sources, and still no outcome data anywhere in this spoke (synthesis) — the taxonomy describes how these curricula are built, not which one works.

The forms are not exclusive (ai-hands-on, 2026-08-22)

This page and its neighbours were written as though a curriculum picks one form. ai-hands-on is all three at once, and the split is visible folder by folder: the notebooks run (executable), 3.Neural-Network(NN) and 4.Transformer ask the reader to build neurons, layers and attention by hand (build-from-scratch), and the README carries 19 outbound GEN AI reference links, 6 books marked "not included" and a framework table (meta-curriculum).

The decay argument above is what makes the hybrid worth naming. Each form rots in its own way, so a curriculum that is all three inherits all three failure modes simultaneously — link rot on the 19 pointers, toolchain drift under the PyTorch notebooks, and API drift under the 5.RAG module that calls a live model. The maintenance bill is roughly additive, which is a cost the pure forms do not carry and one a reader cannot see from the outside.

So treat the three forms as axes a curriculum can score on, not boxes it sits in. Every page in this taxonomy should now say which axis dominates rather than which box applies.

Related

meta-curriculum · nirdiamant-prompt-engineering · ai-engineering-from-scratch · interview-preparation