Guide · read first, in week 0
Read this first, in week 0. Then read the week 1 concept page, and nothing else yet. The glossary is there for whenever a word stops you; the worked examples are for weeks 2 and 3, where the numbers arrive.
One support-desk agent, built in week 0 and changed every week after. It works for a software product that exists in two supported versions, 2 and 3, whose manuals are almost identical in wording and different in substance.
Given a ticket, the finished agent finds out which version the customer runs, answers only from that version's documentation with a citation or declines and hands over to a person, looks up the customer's own record and ticket history, answers questions about the ticket data with numbers it did not invent, and proposes the next action, carrying it out only when a person approves. It proves its own accuracy on questions it never sees, survives an attack hidden in the content it reads, and ships as a container image anyone can run.
At the end you defend every decision in it twice: once with your AI assistant in hand, once without it. Both kinds of round now appear in interviews.
| Wk | What you add | The one lesson | Hours |
|---|---|---|---|
| 0 | Placement, your repository, two free model keys | Start with a working environment and no excuses. | 5 |
| 1 | The model as a component; the first two tools | A model is an unreliable dependency. Its output must fit a schema and its tools must be described exactly. | 8–10 |
| 2 | Retrieval at real scale, filtered by version | The same model gives a different answer when you give it different context. Choosing the context is the job. | 8–10 |
| 3 | Declining, citations, your test set and the gate | Without a test you cannot see, every improvement is a guess. | 9–10 |
| 4 | Action tools with approval; analysis; MCP | Agency and model size are both costs. Never let the model act without approval. | 9–10 |
| 5 | Tracing, cost per ticket, injection defence, masking | Once the system can act, content it reads can attack it. Remove the capability first. | 8–9 |
| 6 | Release, rollback, write-up, recorded demo | Say what it did for the business, and defend it with or without an assistant. | 8–9 |
Weeks 3 and 4 are the heaviest. The total is 55 to 63 hours. Sessions with your mentor fall at the end of weeks 0, 1, 3 and 6.
Week 0, in order. The discovery session sets your route. You make your repository from your route's template. Open it in Codespaces. Get free keys from two different providers and store both as Codespaces secrets. Connect your repository to the programme (the link you were sent); the programme then runs a baseline on your untouched starter.
Then read only two things: this page and weeks/1/CONCEPT.md. Nothing else yet.
When unsure what to do next, run make next. It names the week you are on, the file to open, and
the first thing that is not working.
| Focus on in week 1 | Safe to ignore until its week |
|---|---|
| How a model fails, stalls, costs, varies and changes under you | Chunking, embeddings and retrieval metrics (week 2) |
| Structured output with Pydantic and one repair attempt | Test sets, the judge and faithfulness (week 3) |
| Writing a tool schema a model calls correctly | MCP, approvals, LangGraph and the routing table (week 4) |
The weekly loop: branch, make check, make evidence, decision record, pull request |
Tracing, injection and personal-data masking (week 5) |
| Filling the AI-use log honestly from day one | Release, rollback and the business write-up (week 6) |
Foundations you need from day one: writing and testing a Python service, proposing a change through a pull request, reading code you did not write, and checking what a system or an assistant produced against the source. These are the entry bar; if one is shaky, say so at discovery.
Ideas that only make sense once built: why the context window is an attention budget, what a retrieval metric is really telling you, why a judge must be checked against people, when an agent beats a workflow, and what blast radius means. Do not try to master these from reading. Read the concept page, build the week, then reread it.
Every pull request gets automated checks against a fake model, and from week 2 a hidden evaluation on a fixed reference model. You see its totals, never its questions. Each week ends in a one-page decision record whose numbers must match your evidence file. At sessions the mentor watches you work with your assistant, then asks you to explain your decisions without it. Using an assistant is expected and logged. Undisclosed help is the one thing that fails the programme.