Curriculum
The six tracks, and why they are in this order
Everything ClaudeMaster teaches sits in six tracks, and the order is not a menu — it is a dependency chain. Each track assumes you can already do the one before it, because the questions in it are unanswerable otherwise: you cannot design what goes into a context window before you know what a token is; you cannot make a tool call reliable before you can get a stable output format; you cannot judge whether an agent should exist before you can reason about tools, checks and termination. Read top to bottom the first time. After that, your own record decides — the coach probes your weakest relevant concept, so you can enter anywhere and it will pull you back to whatever is actually missing.
The chain runs: the machine → the input → what it remembers → whether you can trust it → what it can reach → when it acts on its own. The first two tracks are about talking to a model. The middle two are about making it dependable. The last two are about wiring it into systems, which is where most real failures live and why they come last rather than first.
1 · Claude Foundations
Assumes: nothing. Afterwards: you can say what the model is actually doing when it answers, choose between chat, API, Claude Code and Desktop for a given job, and treat tokens and the context window as a budget you spend rather than a limit you bump into.
2 · Prompting & Output Craft
Assumes: foundations — you need the token and surface model before prompt structure means anything. Afterwards: you can get a specific format out of the model reliably instead of by retrying, decide between structured output and tool use on purpose, and recognise a prompt-injection attempt in text you did not write.
3 · Context & Memory
Assumes: prompting — context design is prompt design at scale and over time. Afterwards: you can decide what goes in the window and what stays out, choose compaction over truncation deliberately, hand a fresh session everything it needs, and write a CLAUDE.md that is actually followed.
4 · Reliability, Security & Trust
Assumes: context — you cannot verify an answer whose inputs you do not control. Afterwards: you can build a check that catches a wrong answer before a user sees it, tell a confident answer from a correct one, and state plainly what your setup does with sensitive data.
Open Reliability, Security & Trust
5 · Tools & MCP
Assumes: reliability — a tool hands the model real consequences, so the checks come first. Afterwards: you can draw the boundary of a tool, describe it so the model calls it correctly, pick a transport for an MCP server, and reason about what a tool must never be allowed to do.
6 · Agents & Orchestration
Assumes: everything above — an agent is a loop over tools, context and checks. Afterwards: you can decide where a loop terminates, when to fan work out to subagents and when that only multiplies the mess, and when an agent is the wrong answer entirely.
What a track actually is
A track is a presentation unit: a handful of modules, each of which opens a Socratic dialogue on that topic. What is measured is not the module but the named concepts inside it — the coach judges those per turn, flashcards and study notes are generated from the ones you fumbled, and a cold quiz round then tests them with nobody scaffolding you. One record sits behind all four surfaces, which is why progress carries between them. The full loop is described on how ClaudeMaster works, and the two numbers it produces — the work you put in, and how far you have come on your current level — are explained in how grinds and claudescore are calculated.
Reading material for single concepts lives outside the tracks as practitioner guides, and each track page lists the guides that belong to it.
Where to start
If you are new to Claude, start at Claude Foundations. If you already ship with it, start anywhere and let the first dialogue find the holes — open the Socratic coach and answer one scenario.