Clawdemy Lessons

Editorial brief on multi-task RL versus meta-RL, the three meta-RL families (MAML, RL2, PEARL), and the foundation-model parallel to academic meta-learning.

Editorial brief for the exploration lesson: three strategy families, the easy-vs-hard distinction as the main decision criterion, and the RND breakthrough.

Editorial brief outlining deep RL's four open frontiers, how each maps to the core algorithms, the tensions among them, and the syllabus recap.

What you will learn about Subagents (definition, four benefits, fields, creation paths) and Claude Managed Agents, plus the decision frame between them.

A brief on shipping a Claude app to production: the five disciplines, the Usage and Cost Admin API, and the rollout checklist as a deploy gate.

A quick orientation to the alignment problem: three failure modes (specification, proxy, deceptive) and why robustness and monitoring do not solve it.

Orientation to the safety-engineering lesson: the three transferable tools (nines of reliability, eight safe-design principles, tail events) and prerequisites.

Overview of the deployment-time safety lesson: what robustness and monitoring failures are, their six sub-mechanisms, and the classify-and-defend protocol.

Brief on Hendrycks' four AI risk buckets, the sub-mechanisms and historical analogies inside each, and the classify-and-defend three-step move.

A brief on why correct components yield incorrect AI systems: four complex-systems properties, Perrow normal accidents, and when layered defenses break down.

Overview of how game theory frames multi-agent AI risks: Nash versus Pareto outcomes, three failure modes, and four cooperation mechanisms with their limits.

A guided overview of moral uncertainty in beneficial AI: the three strategies, social welfare functions, fairness criteria, and the link to outer alignment.

What the opening AI safety lesson covers: the four risk categories, the discipline-vs-stance frame, prerequisites, audience, and difficulty.

A preview of the four-layer AI governance lesson: what each layer covers, why compute governance leads, prerequisites, and the placement skill you will build.

What this lesson covers: the six agent patterns, their verbatim definitions, learning outcomes, prerequisites, and the decision tree for picking the right one.

A guided overview of prompt caching, context windows, compaction, and context editing: what you will learn, prerequisites, and how the levers fit.

A brief on Model Context Protocol (MCP): the connector request and response shapes, per-tool configuration, the L4/L5/L6 decision frame, and connector limits.

Roadmap for the agent-loop lesson: the workflow-vs-agent distinction, the 30-line loop, the stop_reason vocabulary, tool_choice modes, and loop disciplines.

A quick brief on Agent Skills (durable on-disk instructions Claude reads on demand) and Claude Code, the agent harness that makes reusable prompts shareable.

Brief overview of your first Claude API call: the request, structured response, stateless multi-turn conversations, and the system parameter.

Overview of the Claude tool-use lesson: what you will learn, prerequisites, where it fits, and the single capability it builds (the four-step loop).

The production Messages API at a glance: when to stream, batch, or do neither; stop_reason dispatch; the HTTP error map; SDK retries; and request_id logging.

A roadmap to the lesson on Anthropic-provided tools: the three categories, the server-tool response shape, the pricing rule, prerequisites, and read time.

Model selection roadmap: the capability the lesson builds, the three Claude families and effort dial covered, prerequisites, and the cost math.

The multimodal AI closer: six unifying threads across the track, what it deferred or never covered, and three trajectories for where the field heads next.