Applied GenAI Curriculum for AI PMs · Core Sequence · Checkpoint

Unit 13 — Capstone (CORE checkpoint)

Pick one real system your team is building or considering, and write the one-pager below. This is produced, not read — proof that the CORE pass gave you an operating model, not trivia.

How to use this

Choose one concrete system — real or seriously considered. Fill each section below in a few sentences; aim for a single page total. The point isn't length, it's whether you can answer at all. Every section maps to specific CORE units; if you get stuck, the "revisit" note tells you exactly where to go back. Your typing stays in this page (nothing is uploaded); use Print / Save PDF to keep a copy.

The system in one line

Units 01–03

What is it, who's it for, and why an AI feature at all? Where does it sit on the risk/opportunity ladder, and is this build or buy? Is AI critical or complementary here?

Can't answer? Revisit Unit 01 (what layer you operate on), Unit 02 (non-determinism), Unit 03 (feasibility screen, build-vs-buy, human-in-the-loop).

Architecture: which pattern, and why not the others?

Units 04–05

Name the approach — prompting, RAG, finetuning, agent, or a workflow combining them — and defend it. Is your gap facts or behavior? Are the steps predictable? Say explicitly why you rejected the alternatives.

See a worked example

System: support assistant. Choice: workflow = routing → RAG → generate. Why: failures are information-based (needs current policy docs) → RAG, not finetuning. Steps are knowable in advance → workflow, not agent. Finetuning rejected (docs change weekly; would ossify). Agent rejected (no unpredictable multi-step planning; compounding-error risk not worth it).

Can't answer? Revisit Unit 04 (workflow vs. agent + the 5-point agent checklist) and Unit 05 (the four-way decision tree, "facts vs. form").

The eval plan

Units 06 · 07 · 09

What does "good" mean here (2–3 criteria)? How is each measured, at which component? What's the usefulness threshold, and how does the key metric map to business impact? Which eval levels (unit tests / human&model / A/B) run at what cadence?

Can't answer? Revisit Unit 06 (why evals, the three levels), Unit 07 (the four criteria buckets, usefulness threshold), Unit 09 (the 3-step pipeline: components → guideline → methods & data).

The production plan

Unit 10 (+08)

Sketch the architecture boxes in order: context, guardrails (in/out), router/gateway, caches, any agent loop or write actions. For each, name the risk it addresses. Is prompting disciplined and versioned?

Can't answer? Revisit Unit 10 (the 5-step architecture build-up + the question per box) and Unit 08 (disciplined, versioned prompting).

Data & cost risks

Units 11–12

What data does this need (quality / coverage / quantity), and what's the annotation budget reality? Where's the latency and cost pressure — and does the token math even allow your target experience? Online or batch?

Can't answer? Revisit Unit 11 (quality/coverage/quantity, annotation budgeting) and Unit 12 (prefill/decode, TTFT/TPOT, the latency floor, online vs. batch).

The one incisive question

Synthesis

Given all of the above, what's the single sharpest question you'd put to your engineers about this system right now — the one whose answer most changes the plan?

Stuck? A good question usually targets the weakest link surfaced above — often the eval number behind a threshold, or a guardrail's blast radius.

How to grade yourself A complete one-pager isn't about polish — it's about coverage. If any section is blank or hand-wavy, you've found a real gap in your operating model, and the "revisit" note is your remediation. Redo the named unit, then finish the section. This is the whole purpose of the checkpoint: it converts "I read it" into "I can actually reason about a system."

What passing looks like

You've cleared the CORE pass when you can:

Justify one architecture pattern and say why the other three are wrong for this system.
State what "good" means, how it's measured per component, and what score makes it worth shipping.
Draw the production architecture and name the risk each box addresses.
Name the data and cost/latency risks concretely — including whether your target experience is physically possible.
Ask the one question that most changes the plan.

That's the operating model. From here the DEPTH units (14–24) are pulled in on demand — the eval thread continues at Unit 14, the highest-ROI unit in the series, which turns the traces your Unit 09 pipeline logs into systematic error analysis.