Evals and the improvement loop

All videos (44)

• FreeVideo 16:34What Claude is and where you can use itAnthropic, the basic idea of an LLM, Claude's key strengths, and all the entry points (apps, Claude Code, API, cloud) that build the same Messages API request.+
• Video 26:50How the request and context window are builtThe API is stateless, so the full context is sent every time; the tools, system, messages order; tokens and the context window.+
• Video 36:31Gateway, model and response: a request's journeyThe request passes the gateway, then the model (tokenizer, transformer, thinking, generation), and the response returns as JSON or an SSE stream.+
• Video 45:29The tool loop and output to the userThe model only writes tool_use and someone else runs the tool, looping until end_turn; then the forms the answer can take for the user.+
• Video 56:13Long conversations and context managementCompaction, context editing and mid-conversation system messages when context fills up, plus an overview of the Safety and Cost rails.+
• Video 65:49Inside the model: how a token is producedA zoom into stage 4: tokenizer, embeddings, attention, probabilities and sampling, with one worked example question.+
• Video 76:59How the model is trained, and why it has limitationsPretraining, RLHF, Constitutional AI and the constitution/character, and the limitations that follow from them (cutoff, hallucination).+
• Video 87:30The Claude model family and how to choose a modelTiers, current models, which model to use when, history, model lifecycle, model cards and benchmarks, and how to answer Claude vs GPT vs Gemini.+
• Video 96:23How claude.ai works internallyA trace of one real request (with web search) and how the app's features are internally a combination of model, context and tools.+
• Video 106:46Plans, privacy, Artifacts and EnterprisePlans, privacy and data, a deep dive on Artifacts and Projects, Claude in other places, and Team/Enterprise admin and compliance.+
• Video 115:52Prompt engineering: core techniquesClear instructions, system prompt, XML tags, few-shot, chain-of-thought, chaining, the structure of a good prompt, and long-context tips.+
• Video 127:19Reducing hallucination, new-model prompting and use-case templatesWays to reduce hallucination, how prompting changed for newer models (what replaces prefill), and support bot, routing and extraction templates.+
• Video 138:13Claude API fundamentalsSetup, your first request, parameters, the response and stop_reason, the multi-turn format, API versioning and beta headers.+
• Video 147:53SDKs: C#, TypeScript and other languagesC# SDK setup and a basic request, the TypeScript SDK, and install steps and examples for Java, Go, Ruby and PHP.+
• Video 156:46Vision, PDF, Files and CitationsImage and PDF input, image tokens, the Files API and citations, with Python and C# code.+
• Video 166:06Extended thinking, adaptive thinking and effortThinking configuration, effort levels, the old budget_tokens approach vs the new adaptive approach, and preserved thinking.+
• Video 176:48Prompt caching: request 1 vs request 2A cache is a prefix match; cache write and read cost, breakpoints, ordering, and a cache-friendly agent loop.+
• Video 188:16Cost: how the bill is built and how to reduce itA worked cost example, the other cost and latency levers, and the pricing of tools and features.+
• Video 197:33Tool use: definition, manual loop and Tool RunnerTool schema, the manual loop, Tool Runner, important concepts, C# tools and IChatClient.+
• Video 208:06Built-in server tools and advanced toolsAnthropic's built-in tools, code execution, the memory tool, programmatic tool calling, tool search, the advisor tool, and smaller features.+
• Video 216:45Computer use: how Claude operates a screenThe same tool loop, where the tool is screenshot plus mouse plus keyboard; setup and risks.+
• Video 227:48MCP: Model Context Protocol basicsMCP client-server architecture, JSON-RPC messages, primitives and transports.+
• Video 238:35Building an MCP server, connecting it, and authAn MCP server in Python and C#, connecting to it, remote auth, and the ecosystem.+
• Video 247:37Agents and agentic patternsWorkflow vs agent, building blocks, ways to build an agent, best practices and frameworks.+
• Video 257:00Claude Agent SDK: deep diveWhat the Agent SDK provides, options, custom tools (in-process MCP), hooks, the multi-turn client and subagents.+
• Video 267:08Multi-agent systems and Managed AgentsWhen and how to use a multi-agent architecture; the core objects and advanced features of Managed Agents.+
• Video 275:30Platform tooling and ecosystemThe ant CLI, admin, monitoring and auth, Claude Console tools, and learning resources.+
• Video 287:21Agent Skills: what they are and how they loadHow skills and plugins load, progressive disclosure, SKILL.md, and the skill lifecycle and priority.+
• Video 297:01Writing your own skill: from frontmatter to descriptionFrontmatter fields, invocation control, arguments, supporting files, a full example, limits, and writing a good description.+
• Video 305:53Plugins: what they are and how they loadWhen you need a plugin, the 3 stages from install to session, how each component loads, cost, install scopes and the plugin ID.+
• Video 316:18Building your own plugin and marketplaceYour first plugin end to end, creating and sharing a marketplace, versions and updates, and a debug checklist.+
• Video 327:30Claude Code: what happens inside a sessionInstall, how it works, a trace of a bug-fix session, CLAUDE.md memory, and the best workflow.+
• Video 336:43Claude Code: commands, permissions and hooksSlash commands, shortcuts, permission modes, hooks, and plugin/MCP integrations.+
• Video 347:37Claude Code: advancedSubagents, customization, memory and history, automation and scheduling, workflows, sandboxing, CLI flags, where it runs, and review skills.+
• Video 356:30Claude Code in CI/CD and automationGitHub Actions setup, headless scripting and a custom status line.+
• Video 367:31RAG: answers from your own documentsIndexing and retrieval, embeddings (Voyage), contextual retrieval and a mini RAG code example.+
• Video 376:07Evals and the improvement loopDecide with numbers, not feelings: eval sets, graders and LLM-as-judge code.+
• Video 385:41Structured outputs: guaranteed valid JSONConstrained decoding, and what is and is not guaranteed.+
• Video 396:47Streaming and the Batch APIThe flow of SSE events, streaming in the SDKs (C# too), and the Batch API: thousands of requests at half the price.+
• Video 408:49Errors, retries and rate limitsThe path for 429/529/5xx errors, retries and timeouts, rate limits and tiers, refusals and fallbacks.+
• Video 417:16Safety and security: defense in depthRSP, safety mechanisms, interpretability, prompt injection and jailbreaks, and layered defense.+
• Video 427:08What to use when: decision tree and system designHow to choose an approach and a model, a comparison cheat-sheet, and a system design scenario.+
• Video 4310:57Interview quick revision: top questions and codeShort answers to the top questions and code questions: a revision of the whole course.+
• Video 448:25A-Z glossary and rapid fireAll the glossary terms, a practice checklist, and 1-line rapid-fire answers.+