One idea, one rule, one trap, one test. Most lessons end on a real timed mock — those are live now, so you can start before the videos land.
01
The One Question — Persuade / Constrain / Enforce ★
The framework the whole course hangs on, taught with a road analogy: a speed-limit sign (persuade), a speed bump (constrain), an engine limiter (enforce). Pick the weakest tier that fully solves it — unless the stem says guaranteed, tamper-proof, cannot be bypassed, must always.
5:30no mock test
02
Start Here — What This Course Is ★
The course-intro piece, and the only lesson with a face to camera throughout. What the exam actually is, how the course maps to it, and the kill list — what to stop studying. Ends by telling you to sit Test 1 cold, before you go further, so you have a baseline to improve on.
03
The Agentic Loop & stop_reason
Taught with a walkie-talkie, not with code: on a radio you never guess whether the other person has finished — they tell you, in one word. “Over” means your turn; “out” means finished. stop_reason is exactly that, and it is the only thing that ends a task. Covers the three anti-patterns the official guide names, the correction to my own notes, and agent vs workflow as a train and a taxi.
04
Subagents, the Task Tool & Partitioning
Taught as a newsroom: an editor, three reporters and a stack of briefs. A reporter knows only what is on the brief — subagents do not inherit context. Hand all the briefs out before anyone leaves — parallel means several Task calls in one response. And a story each, not a stage each: partition where the work is genuinely independent. By phase is a pipeline with barriers wearing a partition's clothes.
8:48no mock test
05
Hooks, Prerequisites & Guarantees
Hooks do two jobs: PostToolUse normalises data before the model sees it, and tool-call interception blocks policy violations. Where money moves, a programmatic prerequisite beats a prompt — prompt-based compliance has a non-zero failure rate.
06
Named Patterns & When Not to Use an Agent
Agent versus workflow — the most-tested idea in Domain 1. If the path is known in advance, it is a workflow. Agents are for deciding the path. Plus routing, hub-and-spoke, and recognising ReAct and plan-execute from a description.
The biggest cluster of lost marks. Glob finds by name, Grep finds by contents. You must Read before you Edit, and Edit needs a unique match. The blocked-or-impatient test: blocked means Write, impatient means incremental Edit.
08
Config Hierarchy, Commands, Skills, Headless
The four config mechanisms and which file wins — narrower scope beats broader. Slash commands, skills and argument-hint, plan mode, --output-format json with --json-schema for CI gating, and resume versus continue.
09
Prompt Engineering & Structured Output
Describe versus show versus constrain: descriptions fix tool selection, examples fix format, schema fixes values. Retry fixes format and structure — retry will never produce information absent from the source.
Tools are model-controlled, resources application-controlled, prompts user-controlled. Around 18 tools badly degrades selection; 4 to 5 is the target. Invalid parameter combinations mean split the tool — make the mistake impossible to express.
11
Context Management & Reliability
Progressive summarisation destroys numbers, dates and amounts — keep a persistent case-facts block outside the summarised history. Trim tool results before they accumulate. A bigger context window does not fix attention quality.
12
Escalation & the Support Scenario
A whole scenario, and the easiest one to under-prepare. Escalate on an explicit request for a human, on genuine policy gaps, or on no meaningful progress. Never on sentiment, never on self-reported confidence.
6:25no mock test
13
Evaluation & Monitoring — the Blind Spot
An aggregate number can be excellent while a segment inside it is catastrophic: 97% overall can be 60% on the tenth document type. When a stem stresses overall or on average, the answer involves disaggregating. Analysis precedes pilot.
14
The Ladder, Assembled ★
The payoff. Nothing new — every mechanism has been taught, so this lesson sorts them into nine numbered rungs and turns the framework into a lookup you can use under time pressure. Rungs 1–4 persuade, 5–6 constrain, 7–9 enforce.
15
Final Review — What the Mocks Taught Me ★
The honest one, built from five timed mocks and the error analysis behind each. Four root causes, three of them mechanical and fully learnable. Plus the keyword-to-answer cheat sheet, read aloud. Ends on the full 60-question, 120-minute mock.