Module 20 · Capstone

End-to-End Engagement Simulation

Estimated time: 90 minutes Track: Yungsten Tech Employee Curriculum · All staff (engineer completion required)

About this capstone

This is not a reading exercise. It's a simulation of the full Yungsten workflow on a realistic client scenario. You will produce ten deliverables — the same artifacts you'd create in a real first-quarter engagement. Each is graded against the model answers at the end.

Complete the deliverables in order. Each builds on the one before.

The scenario: Midcoast Advisory Group

About the client:

Midcoast Advisory Group is a 45-person independent registered investment advisor (RIA) based in Portland, Maine. They manage approximately $1.2B in assets under management across 380 high-net-worth client households.

The intake email you received:

"Hi — we've been watching the AI space and think there's something real here for us. Our senior advisors spend an embarrassing amount of time every week writing client commentary. Every quarter we produce personalized portfolio update letters for all our clients — about 380 of them — and it takes our team of 6 senior advisors roughly 2-3 hours each to write their commentary section. That's 12-18 hours per advisor per quarter just on writing. They hate it. We'd love to get that down. We're a regulated firm (SEC-registered) so we need to be thoughtful about data. Happy to chat whenever works."

— Marcus Chen, COO

Additional context you've gathered:

The portfolio update letter has four sections: market commentary (firm-wide, same for all clients), portfolio performance summary (pulled from their portfolio management system, Orion), advisor commentary (personalized by advisor, 2-4 paragraphs), and next steps (templated).
The advisor commentary section is what's time-consuming. Advisors currently write it from scratch each quarter using a combination of Orion performance data, their meeting notes from the past quarter, and their knowledge of the client's situation.
The firm has a compliance officer (Sarah Kim) who must approve any new technology use.
Their data classification situation: client names, account numbers, and financial data are all considered PII under their compliance policy. They have not previously used any AI tool with client data.
The designated operator for any AI tool would be the advisors themselves — all 6 are senior, non-technical, financially sophisticated professionals aged 40-60.
Their IT infrastructure: macOS laptops for all advisors, Google Workspace, Orion for portfolio management (has an API), no internal engineering team.

Your deliverables

Deliverable 1 — Intake brief (Module 15: Client Intake)

Write the pre-call intake questions you'd send back to Marcus before your first scoping call. Cover all four intake dimensions. Keep it to 8-10 questions.

Deliverable 2 — Engagement rung recommendation (Module 15: 90-Day Model)

Based on what you know, which engagement rung do you recommend (Starter, Full, Custom) and why? What would the first-90-days plan look like at that rung?

Deliverable 3 — Agent specification (Module 16: Named Agent Design)

Scope a single named agent for this engagement's first delivery. Use the one-page specification format:

Agent name:
Primary job:
Inputs:
Outputs:
Constraints:
Operator:
Success criteria:

Deliverable 4 — System prompt (Module 16: System Prompts)

Write the full five-section system prompt for the agent you specified in Deliverable 3. Include all five sections: identity, expertise/context, operating instructions, constraints, output format.

Deliverable 5 — MCP configuration (Module 16: Claude Desktop MCP)

Write the claude_desktop_config.json entries for the MCP servers this agent needs. Include version pins (use placeholder version numbers). Note: you will NOT be connecting to Orion's live API in the initial deployment — advisors will paste relevant data. Think about what the agent actually needs.

Deliverable 6 — Test plan (Module 16: Agent Testing)

Define all five test cases for this agent before delivery. For each: describe the test input, what you expect, and what a pass/fail looks like.

Deliverable 7 — Functional operator runbook (Module 15: Defining Done)

Write the functional runbook for a senior financial advisor who has never used a terminal and is not technical. Cover: what the agent does, how to use it, what success looks like, the three most likely failure modes and fixes, and escalation.

Deliverable 8 — Wiki entry outline (Module 17: Wiki Architecture)

Write the wiki entry for this agent (the /02 - Agents/[Agent Name].md file). Use appropriate markdown structure. Include: overview, agent spec summary, system prompt version log, operating instructions reference, known edge cases, and change history.

Deliverable 9 — CLAUDE.md AI stack section (Module 17: CLAUDE.md)

Write the AI stack summary section that would appear in Midcoast's organizational CLAUDE.md after this agent is deployed. Include: what's deployed, operator, data classification relevant to this agent, and current engagement phase.

Deliverable 10 — Architecture document (Module 19: Technical Handoff)

Write the architecture document for this agent (the engineering handoff artifact). Include: overview, components table with versions, data flow, external dependencies, and known limitations.

Scoring guide

After completing all ten deliverables, compare against the model answers below. For each deliverable, assess:

Full credit: Hits all required elements, demonstrates the principle, would work in a real engagement
Partial credit: Missing 1-2 elements but directionally correct
Insufficient: Missing core elements or misapplies the principle

There are no wrong answers to the scenario — there are better and worse applications of the frameworks. The model answers represent one strong response; yours may differ and still be correct.

Model answers follow in the prompt exercise section.