Course screens as the founder pastes them. He reports nothing new so far, so pitch quizzes above this level. Testable specifics are flagged in bold.
Module 1 - How LLMs behave (tokens, context, sampling, non-determinism)
- Chars-per-token depends on the model's tokenizer; everything (system, history, tools, tool results, output) counts.
- Input larger than the window: validation error before generation. Fits on input but hits the ceiling during generation: stops with partial output, stop_reason
model_context_window_exceeded(not an error). - Newest models reject non-default temperature/top_p/top_k with a 400; steer them by prompting. Temperature 0 is not deterministic.
- Tests assert properties (field present, value in range, parses), not exact text; model-graded evals for meaning.
Module 1 - Model options and reasoning modes
- 4 tiers: Fable (most capable) > Opus > Sonnet (balanced default) > Haiku (speed/cost). Start with Sonnet; move up or down only when an eval justifies it.
- Reasoning is a per-request setting, independent of model. Adaptive thinking +
effort;budget_tokensis deprecated and returns a 400 on the newest generations. - Thinking is omitted from responses by default on the newest models; request summarized display. Per-model thinking defaults differ.
Module 1 screen 4 - Prompting modes (zero / one / multi-shot)
- Examples sit in the prompt (not training data) and show the output shape.
- Each example costs tokens on every call; use the smallest prompt that is reliable.
- One or two examples usually fix a structure/casing/edge-case miss faster than more instructions.
- Examples can let a cheaper model match a bigger one: pick the simplest model plus the fewest examples that meet the eval.
Module 1 screen 5 - Technical substrate (SDKs, REST, streaming, async)
- The SDK (Python, TypeScript, others) is a thin layer over the same REST API: auth, request building, retries, parsing.
- Synchronous for short responses and backend jobs; streaming = server-sent events (SSE) over the same HTTP connection for long or user-facing output. Module 2 covers consuming streams safely and recovering from interruptions.
- Python:
AsyncAnthropicasync client. TypeScript: the standard client is Promise-based; there is NO separate async client class. Async = concurrency without blocking; results are still real-time. - Message Batches: submit many, get an ID, poll; up to 24h; lower per-token cost. For offline pipelines, eval runs, and bulk jobs where cost matters more than turnaround.
- Trap: "async client" (concurrency, real-time) vs "Batches" (offline, cheaper, up to 24h) solve different problems.
Module 1 screen 6 - module quiz (founder 4/4: B, C, B, C)
- Q1 non-determinism (B), Q2 model choice vs reasoning mode (C), Q3 zero-shot already works, so examples add cost (B), Q4 offline + lowest cost = batch (C).
- Calibration: the key was the longest, most specific option every time, and the distractors were throwaways. The real exam length-matches its options, so this quiz overstates his readiness. Our quiz deliberately breaks that tell.
Module 1 screen 7 - predict-the-behavior exercise (founder 4/4; the longest answer won every time again)
- New testable detail: the oldest context is never silently dropped; trimming or summarizing is the application's job. (Distractor: "the model silently drops the oldest turns.")
- A synchronous loop at 50k documents hits rate limits; streaming earns nothing with no one watching.
- Calibration: 8/8 so far on course checks, all solvable by length alone. The course is low-yield for him; use it for coverage and spend prep time on the quiz ([[2026-10-20-ccdv-f-quiz-1]]) and docs.