New: Roadmaps ordered paths through our cheat sheets and flashcards, so you always know what to study next.
Explore themSee what's new on GitHubFrom your first prompt to a shipped, production-grade AI application.
A 15-step learning path. Follow it in order, or jump to what you need.
This path is for developers and career-changers who want to build and ship real LLM-powered applications, not just chat with one. Expect about 8 to 11 weeks at a few hours a week, moving from prompting and API calls through retrieval, agents, and MCP tool connections, into evaluation, safety, and the operations work that keeps an AI system alive in production. It skips training models from scratch and classic ML pipelines, and it leans career-shaped, closing on an interview-facing production build; for the same core skills without that framing, see Generative AI & LLM Engineering. By the end you can call any major LLM API directly, ground answers in your own data with RAG, wire a tool-using agent behind MCP, and ship a production-style AI application for your portfolio.
Expected: comfortable writing and running basic code and calling a REST API (Python or JavaScript is fine). Helpful but not required: prior exposure to machine learning concepts or working with SQL databases.
Start here: reword a single prompt and watch a model's answer change today, no account or setup required, the same instinct you will lean on in every step ahead.
Now see why those wording tweaks worked: tokens, attention, and training decide what a model can and cannot do, the internals step 1 let you skip.
Take the prompting instincts from step 1 and the model behavior from step 2 and wire them into a real call against OpenAI, Claude, or Gemini.
You can shape a prompt that lands and call OpenAI, Claude, or Gemini directly from your own code. Next up: forcing that output into a shape your code can actually trust.
Finish this section to unlock.
+100 XP
Stop parsing messy text out of a reply: force the model's output into JSON your code can trust, the same contract every tool call in step 9 depends on.
This is where prompting stops being about wording and becomes about what the model sees at all, memory, retrieved facts, tool results. Expect to keep tuning this as your app grows, not just get it right once.
Turn text into vectors so similar meaning becomes a distance you can compute, the building block step 7 retrieves against.
Connect those vectors to a live knowledge base and the model finally answers from your data instead of guessing. Most production AI apps live or die on retrieval quality here, not on the model itself.
Your app can retrieve real facts instead of hoping the model remembers them right, and a few due flashcards on chapter 1 keep those API habits fresh while you build on them. Next up: letting the model decide what to do with what it retrieves.
Finish this section to unlock.
+100 XP
If you're aiming at a narrow, high-volume task that RAG cannot fix cheaply, teach the weights themselves instead of feeding fresh context on every call.
Give the model the structured outputs from step 4 and the retrieval from step 7, then let it choose the next tool call instead of you scripting every branch.
Standardize how those agents reach tools and data: one protocol instead of a custom integration for every API you connect.
If you're building an agent that has to see images or hear audio, not just read text, the context habits from step 5 extend to those senses too.
You can hand a model tools through MCP and let it plan its own next step instead of you scripting every branch. Next up: making sure it's reliable enough to hand to a real user.
Finish this section to unlock.
+100 XP
Before you ship any of this, measure whether your RAG answers and agent runs are actually good, not just plausible sounding replies.
An agent that retrieves your data and calls tools on its own is also one that can be tricked into misusing both: close the gaps prompt injection and data leakage open.
Operate what you have built: track cost, latency, and drift the same way step 12 taught you to measure quality.
Pull prompting, retrieval, agents, and evaluation into one production build end to end, the shape hiring teams actually test for.
You can prompt, retrieve, orchestrate agents, evaluate, secure, and operate a real LLM application end to end, the same shape hiring teams actually test for. This badge marks that you built the whole thing, not just the demo.
Finish this section to unlock.
+100 XP
Finish every required step, at least 70% of them genuinely done (not skipped), to earn this badge and 500 XP.