Capstone: Build

The capstone ships a complete customer-facing agent: every module's artifact, integrated, tested, and presented as a launch candidate. Two paths: your own (an assistant your workplace could actually deploy — strongly preferred; sanitize per your data policy) or Harbor Helper completed (extend the course build with a sixth in-scope topic of your choosing and full integration).

1. 1Freeze the brief at v2: incorporate every amendment the course forced (the red-team's crack list, the drill findings, the eval failures). The brief's changelog itself is a submission artifact — it shows design responding to evidence, which is the whole method.
2. 2Integrate the stack in one place: the assembled system prompt (scope, persona, grounding rules, injection bounding, never list, distress override, handoff triggers), the corpus with freshness metadata, the simulated account layer, the output-check pass, and the package generator. One document or one pipeline — either way, one source of truth a reviewer can read top to bottom.
3. 3Close your known gaps deliberately: pick the three worst items from your running failure list and fix them properly (not prompt-patch-and-pray — if retrieval was the cause, fix retrieval). Then full regression: base + attacks + flows + canaries, green in a single run, scorecard attached.
4. 4Build the sixth topic (or your equivalent) end to end as proof of method: scope entry → corpus docs → golden conversations → eval cases → guardrail review → integrated. Document the hours it took — the number that shows the system works is how cheap topic seven has become.
5. 5Draft the launch one-pager for Module 6's final lesson: what it handles, the quality floor and current scores, the handoff story, the incident path, and what stays out of scope at launch and why. Written for the support-team lead and the exec sponsor simultaneously — the two audiences whose trust the launch spends.