ModernCS
Join for free
Claude

Generative AI · Taught with Claude

Generative AI Engineering

The craft under every AI product: streaming, tool calls, structured output, caching, and cost. You learn where the token bill and the latency actually come from, and ship something people can rely on.

Download the course plan
  • Advanced
  • 4 phases
  • 25 sessions
  • First session free

Syllabus

4 phases · 25 sessions · each phase ends in something you have built

1

The Wire

Drive the Messages API directly and account for every field in the request and the response.

$108

this phase

  1. 1.1

    Messages, Not Chat

    Free preview

    One stateless endpoint, the whole history resent every turn, and where the system prompt lives.

    Open
  2. 1.2

    Streaming, Event by Event

    message_start to message_stop, deltas accumulated by hand before you trust the SDK helper.

  3. 1.3

    Every Stop Reason

    end_turn, max_tokens, tool_use, pause_turn, refusal, and the branch most code forgets.

  4. 1.4

    Thinking and Effort

    Adaptive thinking, the effort ladder from low to max, and why budget_tokens now returns 400.

  5. 1.5

    Counting Tokens Honestly

    The count_tokens endpoint, why tiktoken undercounts Claude, and reading usage on every call.

  6. 1.6

    Errors That Retry

    429 and 529 back off, 400 never does, and retry-after beats your own exponential guess.

By the end of this phase:A streaming CLI client with correct stop-reason branching, bounded retries, and per-call token accounting.
2

Data, Not Prose

Make the model return typed data your code consumes without a regex or a hope.

$108

this phase

  1. 2.1

    Structured Outputs

    output_config.format with a JSON schema, and why output_format and assistant prefill are gone.

  2. 2.2

    Schemas That Compile

    The supported JSON Schema subset, additionalProperties false, and the first-request compile cost.

  3. 2.3

    Tool Definitions

    Name, description, input_schema, and descriptions that say when to call, not just what.

  4. 2.4

    Strict Mode

    strict true on the tool, parsing input as JSON, and never matching a serialized argument.

  5. 2.5

    Parallel Tool Calls

    Several tool_use blocks in one turn, and all results returned in a single user message.

  6. 2.6

    When The Shape Breaks

    Truncation, refusals, and the validation error you count instead of silently swallow.

By the end of this phase:An extraction service that validates against a schema and reports its own failure rate per field.
3

Context and Cost

Cut spend and latency on a real workload without moving output quality.

$108

this phase

  1. 3.1

    The Prefix Is The Key

    Render order of tools, system, messages, and how one timestamp invalidates everything after.

  2. 3.2

    Breakpoints and TTLs

    Four cache_control markers, 5 minute against 1 hour, and the per-model minimum that fails quietly.

  3. 3.3

    Auditing a Cache Miss

    cache_read_input_tokens at zero, unsorted JSON keys, and a tool list that varies by user.

  4. 3.4

    Budgeting the Window

    A million tokens is a budget, not a bucket, plus context editing and server-side compaction.

  5. 3.5

    The Batch Lane

    Half price, a 24 hour ceiling, and results keyed by custom_id because order is not promised.

  6. 3.6

    Picking a Model

    Opus 5 against Sonnet 5 and Haiku 4.5, measured on your own task, not on a leaderboard.

  7. 3.7

    The One Second Budget

    Time to first token, prefill against generation, and the 200ms you are never getting back.

By the end of this phase:A before-and-after benchmark reporting cache hit rate, p95 time to first token, and cost per request.
4

Shipping It

Treat the prompt as a versioned artifact and know what every request cost, took, and returned.

$108

this phase

  1. 4.1

    Prompts as Artifacts

    Prompts in git with immutable version ids, pinned per environment, rolled back by pointer.

  2. 4.2

    Regression Before Deploy

    A fixture set, a diff against the last version, and the gate that blocks a bad prompt.

  3. 4.3

    Tracing a Request

    OpenTelemetry spans carrying prompt version, model, tokens, cache hits, and dollars.

  4. 4.4

    Cost Per Success

    Per-call dollars, retries that quietly double spend, and the number that actually matters.

  5. 4.5

    Streaming to a Browser

    SSE over HTTP/2, proxy buffering, reconnects, and a half-written answer left on screen.

  6. 4.6

    The Failure Runbook

    Rate limits, overload, schema drift, and what your feature does when the API is down.

By the end of this phase:A deployed feature with pinned prompt versions, traced requests, and a rollback you have actually performed.

Tools you will use

  • Claude Opus 5
  • Claude Sonnet 5
  • Claude Haiku 4.5
  • anthropic Python SDK 0.120
  • Claude Code
  • Pydantic 2
  • FastAPI
  • OpenTelemetry GenAI semantic conventions
  • Langfuse
  • promptfoo

What you will build

  • Streaming Extraction Service

    Typed JSON from free text, streamed, with a measured schema failure rate

  • The Cost Teardown

    One workload, caching and batch applied, spend and p95 latency cut on record

  • A Feature People Depend On

    Pinned prompt versions, traced requests, and a rollback performed under load