The parts that are still moving
The 425 concepts are settled ground. Backpropagation and eigenvectors will mean the same thing in five years. These will not — they are the craft of building with models rather than the theory beneath them, and practice rewrites them every few months. That is why they sit here and not in the curriculum: a prerequisite graph assumes the thing being learned holds still.
None of these are written yet. They are listed so you know what is coming rather than finding an empty page behind a hopeful link.
Prompt engineering
being writtenWriting the instruction so the model can follow it — and, more often, finding the sentence that was missing when it did something odd.
Context engineering
being writtenDeciding what the model gets to see. Retrieval, summarisation, what to keep across turns and what to let go of, and why a longer window does not settle any of it.
Harness engineering
being writtenThe scaffolding around a model: the tools it can call, the checks on what it returns, and what happens when it fails. Most of the reliability lives here rather than in the model.
Loop engineering
being writtenAgents that keep going until the work is done — how a loop knows it has finished, how it avoids running forever, and what it should do when it is stuck.
Graph engineering
being writtenStructuring knowledge as things and the links between them, so a system can answer questions that no single passage contains.
Web MCP
being writtenThe Model Context Protocol over the web: how an agent reaches tools and data it does not own, and what that opens up and gives away.
Meanwhile the curriculum itself is free to read and downloadable as a skill pack for your own agent.
