AI 101 · Module 2
From prompting to document analysis and Cowork practice: building your own Use Case
Start Section 1You ask Claude to draft a price-inquiry email to a vendor, but the first version sounds too stiff. What's the most efficient next step?
Think of a prompt as handing a task to a colleague: the less context they have, the more clearly you need to explain it. A vague instruction produces a vague result.
Info
Your manager says "take a look at this report," and you paste that sentence directly to AI as the prompt. What is most likely to happen?
You ask AI to write a handover note for a new hire, but you don't say who will read it, and you don't attach the current workflow document. Which two T-C-R-F elements are missing?
Help me analyze the customer issue.
In T-C-R-F, what does "Reference" correspond to?
Document analysis is one of the most common work scenarios: meeting notes, reports, and customer feedback all need to be read, organized, and questioned. Here's a reusable six-step process.
While analyzing a cross-department project meeting note, you find a task with "no owner assigned yet." Which of the six steps does this gap belong to?
Info
A meeting note says "this month's progress is two weeks behind," but gives no reason at all. If the AI summary adds "likely due to understaffing," what category does that sentence belong to, and how should it be labeled?
The source data says "42 complaints last quarter," but the AI summary says "over 50 complaints last quarter." Per evidence-verification principles, what should you do?
Put frequently used templates, style guides, and background material into a Project so every future task continues from the same context.
A customer-service team puts its FAQ templates and brand-voice guidelines into a Project, so every agent replying to customer emails continues from the same tone and format without re-explaining it each time.
A customer-service team handles a large volume of emails daily and must keep a consistent brand voice. What is the most efficient approach?
When a task needs repeated edits before final delivery (a report, a table), Artifacts let you keep editing without regenerating the whole thing.
You're building a project-status table for your manager to review, and it will go through several rounds of feedback. Editing it as a single Artifact makes it far easier to track the revision history than re-pasting the whole table each round.
Using the four practice files (Meeting_A, Meeting_B, Sales_Report, Customer_Feedback), practice delegating a fairly complete task to Claude and producing: three main issues, evidence, open items, recommendations, a one-page executive summary, and a Fact / Analysis / Unknown breakdown.
When delegating a Cowork task, why ask for a Fact / Analysis / Unknown breakdown in the output?
Before delegating a complex task involving four documents that needs to produce an executive summary, what is the best first step under plan-before-execute?
After the output is produced, review it using Human–AI Review principles: can facts be traced back, is the analysis reasonable, were open items missed. Only then deliver the executive summary.
The Cowork executive summary says "customer satisfaction declined across the board," but one of the four source files actually shows satisfaction rising. Catching this contradiction during review means you must have AI correct it or adjust the wording yourself; it cannot ship as-is.
Using Sales_Report.csv, practice finding: missing values, duplicates, completion rate, months that missed target, longest response time, lowest satisfaction score, and possible relationships. Do not claim causation; describe only the observed correlation.
Warning
Analysis of Sales_Report.csv finds that "months with longer customer-service response times also have lower satisfaction scores." What is the most rigorous way to describe this?
Revisit Module 1's decision flow: single-turn tasks use Chat, multi-step multi-source tasks use Cowork, recurring well-defined tasks use Code. Which mode does each of today's four exercises map to?
Which mode best fits analyzing Sales_Report.csv for completion rate and possible correlations?
Pick a task that actually recurs in your own work and fill out the Use Case Canvas: current process, frequency, time spent, pain points, the AI help you want, the suggested mode, and the consequence and human-review method if it goes wrong.
Warning
Before the course ends, fill out the Exit Ticket: the work you most want AI help with, its frequency and time cost, an initial tool category, and what you want the follow-up department interview to cover.
