Sync my progressJoin a session to keep progress across devices and share it with your instructor
Learning objectives
Distinguish generative AI from agentic AI, and name the main limits of generative AI
Explain the Chat, Cowork and Code work modes and when each one fits
Describe how Projects and Artifacts support long-running context and iterated output
Apply the human-review checklist and the four data-safety questions before using AI output
Course Orientation
8 min
Why Learn Claude
This course isn't about memorizing where a button lives. Interfaces change, but knowing which tasks to delegate to AI, how to give clear instructions, and how to review results are durable skills.
Goal: build a transferable AI work mode, not muscle memory for one UI
Scope: three primary work modes (Chat, Cowork, Code)
Principle: Human → AI → Human. A human defines the task and stays accountable for the result
Info
AI provides capability; the human keeps responsibility. This idea runs through the whole course.
Key takeaway: This course teaches work modes and judgment, not a specific UI.
Generative AI
10 min
What Is Generative AI
Generative AI produces new content, such as text, summaries, code, and draft analysis, from what you give it: text, documents, or images. Its core ability is generating a plausible continuation from context.
Common uses
Drafting, summarizing long documents, translation, meeting-note cleanup, code snippets.
Admin work: turning a rough event-plan draft into a polished version for your manager
Customer service: turning a messy pile of customer emails into a bulleted list of issues
Project work: summarizing a long English technical document into three key points in your working language, for quick team review
Info
Generative AI never decides for itself that a draft is "ready to ship." That judgment call is always a human one.
Check yourself
A colleague asks Claude to "turn this pile of scattered meeting-audio transcripts into a bulleted to-do list." Which core generative-AI ability does this mainly use?
Key takeaway: Generative AI is good at producing new content, but it does not inherently know right from wrong.
8 min
Limits of Generative AI
It can produce plausible-sounding but incorrect content (hallucination)
It does not verify sources unless explicitly asked and given evidence
Its grasp of recent or internal information depends on the context you provide
It can't take real-world action on its own unless paired with tools or agentic features
Warning
Any output needs human review of key facts and numbers before it ships.
Check yourself
What is generative AI's most important limitation?
Key takeaway: Generative AI's output needs human verification, especially for facts and numbers.
Agentic AI
8 min
What Is Agentic AI
Agentic AI does more than generate content. It plans steps, calls tools, executes multi-step tasks, and adjusts based on intermediate results. Claude's Cowork and Code capabilities fall into this category.
It plans "what first, what next" instead of just returning a block of text
It can read multiple files, search for data, run calculations, and stitch the results together
If it finds mid-task that the data is insufficient or the direction is off, it can adjust the next step instead of stalling
Work scenario
Asking Claude to read five branch operating reports for the month, roll up the common anomalies, and produce a summary table requires planning, reading multiple sources, and synthesis: an agentic task, not a single block of generated text.
Check yourself
Which task best demonstrates "agentic AI" rather than plain "generative AI"?
Key takeaway: Agentic AI plans and executes multiple steps, not just a single reply.
8 min
Generative vs Agentic
Generative AI
Single-turn response
You provide input, it produces output
Does not take action on its own
Best for: drafts, summaries, translation
Agentic AI
Multi-step planning and execution
Can call tools, read/write files
Adjusts mid-task based on results
Best for: cross-document work, automation
Same task, two approaches
"Write a letter asking a client for missing documents" usually needs only a single generative pass. But "check whether 20 clients have submitted complete documents, and draft a reminder for each one that is missing something" needs agentic capability to check each one and produce individual drafts.
Key takeaway: Choose generative or agentic based on whether the task needs autonomous multi-step execution.
8 min
AI Autonomy and Governance
More agentic capability means more autonomy, and more need for governance. Before delegating a task, decide the scope of authority, what actions are allowed, and who reviews the result.
Authorization scope must be explicit: what AI can and cannot do
High-impact outputs (external documents, decision inputs) always need human review
Keep a traceable record for later inspection
Work scenario
Letting AI "draft" a batch of billing-adjustment notices for clients is a very different authorization scope from letting AI "send" them directly: in the first case a human still approves the final version before it goes out, while in the second, AI effectively gets the authority to make an external commitment.
Warning
Never design a workflow that implies "AI output can ship as-is."
Check yourself
For an agentic task that compiles and drafts a batch of client billing-adjustment notices, which authorization design best fits governance principles?
Key takeaway: More autonomy demands stronger governance and review.
Claude Work Modes
8 min
Claude's Working Mental Model
Think of Claude as three work modes: Chat (thinking together), Cowork (delegating multi-step tasks), and Code (automation). All three share the same judgment and review principles.
Comparing the three work modes
Mode
Core action
Work example
Chat
Single or multi-turn conversation; thinking and drafting
Brainstorming three directions for a new project
Cowork
Delegating a fairly complete, multi-step task
Rolling up four client feedback files into an executive summary
Code
Repetitive, well-defined automation
Cleaning up the same weekly sales-report format
Key takeaway: Chat, Cowork, and Code are three expressions of the same work mode.
6 min
Chat: Think, Ask, Draft
Chat is best for quick discussion, brainstorming, drafting, and clarifying a problem, in a single turn or a short back-and-forth.
Work scenario
Your manager asks on the spot for three possible event names. Chat lets you brainstorm and filter back and forth, narrowing down to a few candidates in minutes with no files to prepare.
Check yourself
Which scenario is best suited to plain Chat rather than delegating a Cowork task?
Key takeaway: Chat suits single-turn or short multi-turn thinking and drafting work.
6 min
Cowork: Delegate and Execute
Cowork is best for delegating a fairly complete task. Claude reads multiple sources, plans the steps, and produces a result for you to review and adjust.
Work scenario
At month-end you need to roll up weekly performance reports from three regions, find the shared causes behind underperformance, and write an executive summary. This kind of multi-source, synthesize-and-produce-structured-output task suits Cowork, rather than pasting data into Chat sentence by sentence.
Check yourself
You need to roll up three regional weekly reports, find a shared cause of underperformance, and produce an executive summary. Which work mode fits best?
Key takeaway: Cowork suits multi-step tasks that pull from several sources.
6 min
Code: Build and Automate
Code is best for repetitive, well-defined tasks, such as batch data cleanup or generating a report script. A one-off task rarely needs code; a recurring task is worth automating.
Check yourself
When is "Code / automation" the better choice over plain Chat?
Key takeaway: Only recurring, well-defined tasks are worth automating.
Projects & Artifacts
6 min
Projects: Centralizing Long-Running Context
Projects let you gather related documents, instructions, and style guidance in one place, so every conversation continues from the same context without re-pasting material.
Work scenario
You write a weekly or monthly report in the same company format every month. Put the format spec and past examples into a Project once, and after that you only need to paste in new numbers instead of restating the format requirements each time.
Check yourself
You write a monthly report in the same company format every month. What is the most efficient approach?
Key takeaway: Projects centralize context you reuse across conversations.
6 min
Artifacts: Editable, Reusable Output
Artifacts are standalone, continuously editable outputs, such as a document, a table, or a simple web page, suited to content that needs iteration before it is delivered.
Work scenario
Producing a proposal table for your manager often takes several rounds of adjusting fields and wording. Editing it as an Artifact updates that one document in place each time, instead of regenerating the whole thing in the chat and copy-pasting it over again.
Key takeaway: Artifacts suit content that needs iteration and final delivery.
Tool Selection
8 min
Tool Selection: Chat, Cowork, or Code?
1Is the task single-turn and easy to state up front? Yes → lean toward Chat
2Does it need to combine multiple sources and steps? Yes → lean toward Cowork
3Will it recur with well-defined rules? Yes → lean toward Code / automation
4Is the task high-risk or high-impact? Yes → whichever mode you pick, add stronger human review
Check yourself
A task recurs every quarter with fixed rules, but its output directly affects department budget approval. Following the tool-selection flow, what is the best approach?
Key takeaway: Use "single-turn vs multi-step vs recurring rules" as a quick decision guide.
Human Review & Data Safety
8 min
Human–AI Review, in Practice
Separate Fact, Analysis, and Unknown
Trace key numbers and conclusions back to the source data
Check for logical leaps or unverified inferences
Confirm the output still matches the original task goal
Check yourself
An AI drafts: "Complaints rose 12% this quarter, likely because staff are still new to the updated shipping process." Which category does "likely because..." belong to?
Key takeaway: Fact / Analysis / Unknown is the simplest way to review output.
8 min
Four Data-Safety Questions
1Can this data leave internal systems? Check data classification and company policy
2Does it contain customer PII or confidential material? If so, de-identify first or get authorization
3Who can see this conversation? Confirm sharing scope and retention
4What's the cost of getting it wrong? Assess risk level and set the review bar accordingly
Warning
If you're unsure whether data can be used, ask your manager or security contact. Don't decide alone.
Check yourself
A colleague wants to paste a complaint list containing customer names and phone numbers into an AI tool to have it categorized. Following the four data-safety questions, what is the first thing to do?
Key takeaway: Before using AI, confirm the data is allowed to leave internal systems.
AI Workforce Foundation: Six Core Competencies
10 min
What Actually Counts as Being Able to Use AI?
Info
Knowing how to operate a tool is not the same as having AI workforce capability. That capability means correctly choosing the task, supplying necessary context, designing clear instructions, verifying output, keeping human accountability, and knowing when to stop or escalate.
AI Task Judgment: judge whether a task suits AI assistance, distinguish delegable steps from ones needing joint work or sole human accountability, and recognize that high-impact or irreversible decisions must not be left to AI alone
Instruction Design: describe work using Task, Context, Requirements/Constraints, and Output Format, and tell padding an instruction with words apart from supplying necessary work conditions
Context & Source Awareness: judge what information a task needs, distinguish authoritative, supporting, and unknown sources, and avoid unnecessary exposure of personal or sensitive data
Output Verification: check factual accuracy, calculations and quantities, source support, omissions, contradictions, plausibility, and unsupported inference
Human-AI-Human Workflow: a human defines the task, AI assists in executing it, and a human reviews, decides, and keeps final accountability
Responsible Use: recognize personal data, confidential information, unauthorized sensitive content, bias/fairness concerns, high-impact decisions, and situations that need escalation
Check yourself
Which of the following best defines "having AI workforce capability"?
Key takeaway: These six foundation competencies are what this course builds from start to finish, not tool-operation tricks.
8 min
Verification Before Delivery
Before any AI output ships, walk it through a verification checklist systematically -- do not send it out just because it reads plausibly.
1Source check Can this statement be traced back to the source data?
2Number / calculation check Are the numbers, ratios, and totals correct? Does the denominator or baseline actually exist?
3Missing-information check Is an important constraint, exception, or open item left out?
4Contradiction check Do earlier and later parts of the output contradict each other?
5Decision-impact check If this is wrong, how bad are the consequences? The higher the stakes, the stronger the verification should be.
Example
An AI summary says "complaints rose at least 30%, reflecting seriously inadequate training," but the source data only has this quarter's number -- no prior-year figure and nothing about training. "Rose at least 30%" is an unsourced number, and "seriously inadequate training" is an inference beyond the source; both must be flagged for review, never accepted as-is.
Key takeaway: Verification before delivery is the most basic, and most easily skipped, step of output verification.
8 min
Sensitive Data Decision: Can This Go to AI?
Before handing work data to AI, judge which category it falls into, then decide: use it directly, de-identify it first, get authorization first, or never send it to the tool at all.
Judging the sensitivity of common work data
Data type
Recommended approach
Client list (name, phone, purchase history)
Remove or replace identifying fields; keep only what the task actually needs
Student records (grades, attendance, student ID)
De-identify and confirm the use is authorized; keep a human review step
HR records (performance, salary, review notes)
Never use a general-purpose AI tool without authorization; confirm an approved channel first
Assess whether it contains confidential content; strip sensitive passages if needed
Confidential contract terms (amounts, clauses)
Never paste into an unapproved tool; escalate to a manager or security contact if unsure
Warning
If you're unsure whether a piece of data can be used, escalate to your manager or security contact first -- don't assume it's fine on your own.
Check yourself
A colleague wants to paste a document with confidential contract terms into an unapproved public AI tool and ask it to "polish the wording." What is the correct next step?
Key takeaway: Judging whether data can be used at all is the most basic first step of responsible use.
Recap
6 min
Module 1 Recap
Generative AI produces content; agentic AI plans and executes multi-step tasks
Chat / Cowork / Code map to different task shapes
Projects / Artifacts help manage context and reusable output
Human → AI → Human: a human defines the task, AI executes, a human reviews and stays accountable
AI Workforce Foundation, six core competencies: task judgment, instruction design, context and source awareness, output verification, Human-AI-Human, responsible use
Key takeaway: You're ready for Module 2: hands-on practice.
