Course created by A. Karpathy, January 2025

Contextual Design for Solo Entrepreneurs

Your AI agent costs more than you budgeted and the answers are not reliable? Often the problem is not the model but what you put in the context window, when, and at what cost. This mini-course exists to teach those decisions - but the lessons are not written yet: what follows is the syllabus, not the course.

5modules planned
0online
5still to write
ITlanguage

Because today, because this course

  • The cost is seen late.Each token in the context is charged per call. Those who don't measure what goes into the prompt find the bill at month's end.
  • The frameworks change, the principles remain.LangGraph, CrewAI and AutoGen can change APIs or fall behind. Retrieve, enhance, redirect, and observe context no: the course teaches these principles and uses frameworks only as examples.

The syllabus, module by module

  1. 01

    System prompting: cosa dire al modello (e cosa no) In preparation

    Fondamenti di system prompt: ruolo, tono, constraint, policy di rifiuto. Pattern "role + task + context + format + tone".

    ⏱ 45 min
    Not available yet
  2. 02

    RAG strategy: retrieve, augment, generate, evaluate In preparation

    Quando serve RAG (e quando no). Chunking strategies. Embedding model selection. Cross-encoder re-ranking. Evaluation loop.

    ⏱ 75 min
    Not available yet
  3. 03

    Agent orchestration: quando serve un agent (quando no) In preparation

    Architetture multi-agent, LangGraph/CrewAI/AutoGen. Planner / Executor / Reflector. Cost-awareness vs gadget.

    ⏱ 75 min
    Not available yet
  4. 04

    Cost optimization: routing model + prompt caching In preparation

    Routing Haiku/Sonnet/Opus. Prompt caching Anthropic: up to 90% discount on tokens read from cache. Batch API. Streaming vs batch. Token budget.

    ⏱ 60 min
    Not available yet
  5. 05

    Production monitoring: tracing, evals, feedback loop In preparation

    Langfuse / Arize / Helicone / Phoenix. Trace spans, latency, cost per request, hallucination detection. Feedback → prompt iteration loop.

    ⏱ 75 min
    Not available yet

Not applicable

If you are not yet running a model inside a product or a process, this course will be of little use to you. It is not a course about one framework: frameworks show up only as examples. And if you are looking for lessons to read today, this is not the place: for now there is only the syllabus.

Template Editor (Beta)

Coming soon is a tool that generates custom role-based context patterns for PMs, HRS, Legal and Customer Success teams based on your requirements. Currently available as an Open-Source Starter Set on GitHub.

The repository is not public yet: when it is, find it on GitHub.