Advanced Prompt Engineering Capstone Exercise — Advanced Prompt Engineering

What you'll learn

Combine at least three advanced techniques from this course into a single, coherent workflow against one realistic, high-complexity professional task
Produce a defensible, auditable deliverable whose reasoning you could explain and justify to a skeptical reviewer
Self-assess your workflow against the specific failure modes this course has documented — reasoning theater, branch anchoring, agent role bleed, and prompt library rot — and confirm you have guarded against each one you used

Every technique in this course solves a real gap that shows up only on complex, high-stakes work: chain-of-thought verification catches errors basic reasoning misses, tree-of-thought reasoning surfaces options a single line of reasoning would skip, meta-prompting improves the prompts before you rely on them, structured frameworks make reasoning auditable against a standard, multi-agent orchestration catches what a single pass misses, and systematic testing proves a prompt works before you trust it broadly. This capstone asks you to combine several of these techniques the way you actually would on a real, complex professional deliverable — not as isolated exercises, but as one coherent workflow applied to one task with real stakes.

Capstone Exercise

Build a Defensible Strategic Recommendation Using a Combined Advanced Workflow

Context

You are a senior advisor — consultant, analyst, or internal strategy lead, whichever role best matches your actual work — asked to produce a recommendation on a complex decision with at least three genuinely viable options and real consequences if the recommendation is wrong: a market entry decision, a vendor or platform selection, a build-versus-buy call, a restructuring option, or an equivalent high-stakes decision from your own field. The recommendation will be presented to a senior stakeholder who will ask you to defend your reasoning in detail, and it needs to hold up under that scrutiny.

Your Task

Complete the following, choosing a real or realistic scenario from your own professional domain: (1) Use tree-of-thought reasoning to develop at least three genuinely independent options as separate, fully-reasoned branches — in separate conversations, with no cross-contamination between branches. (2) Apply a structured reasoning framework appropriate to the decision type — a weighted risk matrix, a pre-mortem, or an equivalent framework from your field — to evaluate the branches against explicit, named criteria. (3) Run a chain-of-thought verification pass on your leading recommendation, using the checklist technique from Lesson 2, to identify unverified figures and unstated assumptions before finalizing. (4) Optionally, and if the task warrants it, add a critic pass with an explicitly adversarial mandate to challenge your leading recommendation before you finalize it. Write a final recommendation memo of 500-800 words that names your chosen option, the single most important assumption it depends on, and the specific evidence or reasoning that would change your recommendation if it turned out to be wrong.

Your notes (optional)

Deliverable

A 500-800 word recommendation memo naming your chosen option, built on at least three independently-developed branches evaluated against a named structured framework, refined through a documented chain-of-thought verification pass, stating the single most important assumption the recommendation depends on and the specific evidence that would change it. The memo should be something you could defend, point by point, to a skeptical senior stakeholder.

Quick check

The capstone memo must name the single most important assumption the recommendation rests on, and the evidence that would change it. What does this lesson say that requirement is for?

Select one answer.

Key takeaways

A defensible, high-stakes recommendation is rarely the product of a single technique — it typically combines independently-developed options, a structured evaluation framework, and a dedicated verification pass, applied in sequence rather than all at once in a single prompt.
The specific failure modes this course has named — reasoning theater, branch anchoring, framework compliance theater, agent role bleed, test-set overfitting, and prompt library rot — are not abstract warnings; they are the concrete things to check for in your own completed workflow before you trust its output.
Naming the single most important assumption behind a recommendation, and what evidence would change it, is what turns AI-assisted analysis into something a skeptical reviewer can actually engage with, rather than a confident conclusion with no visible seams.
The techniques in this course cost real time relative to a single prompt — that cost is precisely why they are reserved for decisions where getting it right matters enough to justify it, a judgment call you now have the tools to make deliberately rather than by default.