Weeks 23–32, about ten hours a week. Read selected lessons rather than every resource. Weeks 27–28 set aside about 20 hours for the two-week product sprint. Its days are work sessions, not fourteen full working days. Reuse the cases, pipeline and deployment from weeks 23-26.
| Week | Focus | Do this | Done when |
|---|---|---|---|
| 23 | Model as dependency | Selected API lessons; one call with structured output and logging | Model/version, prompt ID, usage, errors and latency recorded |
| 24 | Context + tools | One read-only tool; compare context variants; test malformed arguments | Tool schema and application permission checks in a trace |
| 25 | Retrieval | Label 20 query families; compare keyword, dense and one hybrid/reranked variant | Relevant-source recall and latency reported |
| 26 | Eval workflow | 50 cases; separate development/release families; read 30 development traces | Failure modes and release criteria written down |
| 27–28 | Two-week product | Follow the session plan below | Usable URL + quality/cost report + known failures |
| 29 | Agents vs workflows | Build a bounded loop or explicit workflow; test recovery | Deadline, step limit, permissions and retry behaviour demonstrated |
| 30 | MCP or cost | Current MCP revision or measured caching/routing comparison | Compatible versions or a quality-preserving cost result |
| 31 | Fine-tuning decision | Compare prompt-only, examples and retrieval; optionally an adapter | Keep the simplest acceptable approach; report a null result honestly |
| 32 | Release + buffer | CI regression checks; independent release run; design story | Evidence supports release or a decision to defer |
New words in this chapter (context engineering, retrieval, gold set, pass rate, trace, judge) are explained where they first appear below, and collected in Terms in plain English.