The operator's
curriculum.

Claude is a large language model: a system trained to predict the next piece of text based on everything it has seen before. When you give it a prompt, it is not consulting a live database or "thinking" in a human sense; it is extending the text in front of it in a way that has been statistically useful across millions of examples. Because of this, Claude is extremely good at reading, writing, and transforming text, but it is not a guaranteed source of truth.

You can think of Claude as a text-native coworker. It can read long documents and summarize them, take messy notes and turn them into a clear plan, rewrite copy for new audiences, and even write or review code. It can also orchestrate workflows across tools—like QuickBooks or HubSpot—when connected through skills and integrations. However, Claude will sometimes generate fluent but incorrect answers, especially when asked for specific facts it has no direct access to, or when the prompt is vague or contradictory. That is why you will always pair Claude with clear instructions, guardrails, and human review in this course.

To get consistent value from Claude, you will practice four core habits: delegation, description, discernment, and diligence. Delegation means treating Claude like a junior colleague. You tell it who it is in this task (its role), what it must do (its job), and what success looks like. Instead of writing "Help me with our sales", you will write instructions such as "You are a sales operations analyst; review these ten recent lost deals and find patterns we can act on in the next 30 days."

Description is about giving Claude the context it needs: your product, your customers, your constraints, and good examples. The model has seen a lot of the internet, but it has never seen your exact business unless you show it. Discernment is the discipline of reading Claude's answers critically—checking numbers, comparing to your systems, and asking for clarifications when something feels off. Diligence means turning one-off successful prompts into shared templates and workflows your whole team can reuse.

Not all uses of Claude are equally risky. In this course, you will learn to classify workflows into three groups. Low-risk uses include drafting internal emails, summarizing documents, or brainstorming marketing ideas. If Claude is slightly wrong here, you will catch it easily and the cost is low. Medium-risk uses include drafting customer-facing messages or summarizing analytics; these require human review before anything is sent or used. High-risk uses include moving money, changing prices, deciding on HR actions, or interpreting legal obligations; in these cases Claude is a helper only, and humans make the final decisions.

Trust in Claude grows over time, as you see how it behaves on your own data and as you build evaluation checks around it. You will start by automating low-risk, high-friction tasks like status emails or weekly summaries. As you collect examples, improve prompts, and see where the model fails, you can move selected workflows into the medium-risk category with stronger guardrails. High-risk areas will always keep multiple human checkpoints, and in some domains (legal, medical, regulated finance) Claude's role will remain "summarize and explain" rather than "decide".

Claude's answers often sound confident and polished. That does not mean they are always correct. To work safely, you will learn a simple way to review outputs. First, you break an answer into separate claims—usually one per sentence or bullet. Then, for each claim, you ask: "Do I already know this is true? Can I see where it came from in the input? If I do not know, how would I check?" For important decisions, you cross-check key claims in your own systems or with trusted external sources.

You can also ask Claude to show its thinking. For example, you might say, "List any parts of your answer you are uncertain about" or "Explain which numbers came from the data I gave you and which are estimates." These prompts help you see where the model might be stretching beyond what you provided. Over time, you will build small checklists for common workflows: for a finance summary, you might always verify that totals match QuickBooks; for a support answer, you might ensure that links and policy references exist in your knowledge base.

Typing ad-hoc questions into Claude can be useful, but it does not scale. A workflow is a repeatable pattern with a clear trigger, inputs, steps, and outputs. For example, instead of "Every Friday I ask Claude to summarize our pipeline", you will define a workflow: every Friday at 16:00, pull the latest CRM export, ask Claude to categorize deals by stage and risk, and send a structured summary to a specific Slack channel for review.

Workflows are powerful because they give consistent results across weeks and across people. Once you have the pattern, you can improve it: adjust the prompt, add a checkpoint, or feed in more context. In this course you will repeatedly convert loose prompts into workflows by naming the process, defining inputs (which systems, which files), specifying the structure of the output, and deciding where human approval is required.

You will now pick a single recurring task—often something you do every week—that Claude can help with. Common examples include preparing a short cash note, summarizing key email threads for your team, or turning meeting notes into a clear list of decisions and next steps. The goal is not to fully automate it on day one, but to let Claude do the heavy lifting while you review and edit.

You start by writing down the current process: when you do it, how long it takes, and what the final result must look like. Then you design a prompt that tells Claude your role, your context, and the exact format of the output. You run it once, note what you had to fix, and then adjust the prompt. After two or three iterations, most teams see meaningful time savings. This "reclaim" becomes your first living example of a safe, useful Claude workflow.

In business, you rarely want a vague brainstorm; you want work that is nearly ready to ship. To get that from Claude, you will learn to write briefs that look a lot like what you would send to a human freelancer. A strong brief names the audience, the goal, the business context, the constraints, and the exact structure of the output. It also includes examples of what "good" looks like so Claude can align with your standards.

For example, instead of writing "Write a sales email", you might say: "Write a follow-up email to a prospect who attended our webinar on invoice automation but has not responded for a week. The goal is to re-open the conversation without pressure. Use a friendly, concise tone, reference one specific value point from the webinar, and end with a clear question." When you consistently give this level of detail, Claude's outputs are much closer to ready-to-send drafts.

Your brand has a voice, even if you have never written it down. Claude will mirror whatever voice you describe and demonstrate. A voice document is where you capture this: your tone (formal or casual), your stance (humble or bold), words and phrases you like to use, and those you avoid. You can also include positive and negative examples—"this is us" and "this is not us".

In practice, you will paste or attach your voice doc at the start of a Claude session and instruct the model to follow it. Over time, you can refine the document as you see what works and where the AI still drifts. When your whole team uses the same voice doc, customer-facing content feels consistent, even when it is written by different people using AI.

Claude responds differently depending on the role you assign to it. If you simply say "help me", you get generic help. If you say "You are a senior customer support lead who cares deeply about clear, empathetic communication and accurate use of our policies", the answers change. In this unit you will design persona prompts for key parts of your business: finance, marketing, sales, support, HR, and operations.

Good persona prompts include the job title, level of experience, domain focus, priorities, and things the persona will not do. You will store these personas in a shared library so that everyone on the team can reuse them. This reduces the random variation in prompts between people and makes Claude feel like a stable set of specialist coworkers.

Many of the most powerful workflows involve Claude producing structured outputs that other tools can consume. If you tell the model exactly which fields you want and in what format, you can capture outputs as JSON or tables and feed them into spreadsheets, CRMs, or dashboards. You will practice defining simple schemas and writing prompts that reliably produce data that can be parsed.

For example, after a sales call you might ask Claude to produce a JSON array of "moments", each with a timestamp, speaker, type (pain, objection, pricing, or next_step), and summary. When you do this consistently, you can analyze patterns across calls without manually tagging everything. You will learn how to demand "only valid JSON" or a properly formatted table, and what to do when the model still returns something malformed.

Claude learns a lot from the examples you give it in the prompt. When you provide two or three "before and after" pairs, you are effectively showing the model how you want it to transform inputs into outputs. In this unit you will mine your own archive—great emails, proposals, support replies, and reports—and use them as teaching examples.

You will paste the raw input and the finished output for each example, separated clearly, and then give Claude a new input and ask for a matching output. This few-shot technique is especially useful when your style is distinctive or when you care about structure and phrasing more than pure content. It is often faster and cheaper than training a custom model, and it keeps humans directly in control of what "good" looks like.

As you move from experiments to real workflows, you need a way to know whether Claude is performing well enough. Evals are simple tests you can run regularly. You collect a small set of representative inputs, define what "acceptable" looks like, and then score Claude's outputs against that standard. Over time, you can see whether changes to prompts, context, or models are making things better or worse.

In this course you will create at least one eval set for a workflow that matters to you. For example, if you are using Claude to create weekly KPI summaries, you might keep a set of old weeks where you know what a good summary looks like. Whenever you change the prompt or the model is upgraded, you run the eval again and compare. Evals turn subjective feelings into visible trends so you can make informed decisions about where to invest.

Modern Claude models can read long inputs: full documents, email threads, meeting transcripts, and more. This is powerful, but more context is not always better. If you paste in everything, including irrelevant or conflicting material, the model can get confused—or spend a lot of tokens and time with little benefit. In this unit you will learn to prepare context in layers.

First, you write clear instructions that tell Claude what to do and what to prioritize. Second, you select a small set of authoritative documents—policies, SOPs, reference guides—that it should follow when in doubt. Third, you add task-specific data, such as the particular ticket or contract you are working on. When you are deliberate about what you include and how you label it, Claude can use long context to stay accurate and grounded instead of wandering.

Claude can talk about numbers even when it does not have real data. If you say "What is our revenue this month?" without giving it any actual figures, it will often guess based on patterns it has seen elsewhere. This is dangerous in business. In this unit you will learn to design prompts and workflows that keep numbers grounded and transparent.

You will habitually tell Claude to use only the numbers you provide, to show its calculations step by step, and to label assumptions clearly. When you want estimates—for example, in planning—you will ask for ranges and scenarios rather than single precise numbers. For critical metrics, you will pair Claude with direct queries to your financial or analytics systems so that all figures can be checked. Over time, this "numerical paranoia" becomes one of your most important safety habits.

As your organization starts to depend on Claude, your prompts become important assets. If someone quietly changes a prompt that generates customer emails or compliance language, the effects can be large and hard to trace. That is why you will treat prompts like code: they have owners, versions, and change history.

In practice, this means storing prompts in a shared place—such as a Notion page, internal Git repo, or Atlas skills catalog—along with a short description of their purpose, inputs, outputs, and known limitations. When you make a change, you note what you changed and why. For significant workflows, you might also run evals before and after. This discipline lets you improve prompts over time without losing control or trust.

Accounts receivable (AR) workflows are repetitive and rule-based, which makes them good candidates for Claude support. In an AR autopilot setup, Claude reads your AR aging report and customer notes, groups invoices by how long they have been outstanding and by customer type, and proposes next actions. It might suggest gentle reminders for customers who are only slightly overdue, firmer follow-ups for older invoices, and internal reviews for disputed amounts.

Claude can also draft the actual messages you send. You will provide clear templates and tone guidelines so that outreach remains on-brand and respectful. In the early stages, humans review all drafts before they go out. Later, you may decide that low-risk reminders for small amounts can be sent automatically while larger or more sensitive cases remain human-only. This gradual approach lets you increase efficiency without sacrificing control or relationships.

Leaders need to know the cash situation quickly and clearly. Claude can turn raw bank balances, receivables, payables, and forecasted expenses into a short briefing that anyone can understand. In this unit you will design a standard structure for that briefing, such as: current cash and runway, major movements since last week, a simple 4–8 week projection, and key risks or questions.

You will feed Claude real or sample data and ask it to produce this briefing in plain language, avoiding jargon. Finance team members then check the numbers against the source systems and adjust the narrative as needed. Over time, you can automate the data gathering and send the briefing to a Slack channel or email list every week. The combination of fast AI drafting and human numeric validation gives you a reliable rhythm for financial awareness.

Marketing campaigns often need many pieces: emails, social posts, landing pages, and ad copy. Instead of writing each from scratch, you will learn to feed Claude a single well-structured campaign brief and generate all these assets in one flow. The brief will include the audience, the offer, the core message, the channels you care about, and the tone you want.

Claude will then generate channel-specific versions: a short LinkedIn post, a longer email, a set of headline ideas, and so on. You will check each asset for accuracy, alignment with your voice doc, and compliance with any legal guidelines. The goal is not to eliminate marketers, but to let them spend more time on strategy, testing, and creative direction and less time on first drafts.

After a campaign, you want to understand what happened and what to change next time. Claude can help by turning raw metrics and notes into a structured post-mortem. You will define a consistent template: objectives, overall results, channel breakdown, creative learnings, audience insights, and recommendations.

You then provide Claude with performance data and any qualitative observations from the team. The model drafts the narrative sections, highlighting patterns such as "email opened well but click-through was low" or "paid social underperformed in this segment." Humans check that the numbers match the reports and adjust interpretations as needed. Over time, your library of post-mortems becomes training material for future prompts and strategy.

After sales calls, reps often have to summarize the conversation, update the CRM, and write a follow-up email. Claude can streamline this. When given a call transcript or detailed notes, it can extract the prospect's situation, pain points, objections, timeline, and agreed next steps, then write a clear summary and a personalized follow-up.

You will design prompts that ask for a structured summary and an email in your sales voice. Reps remain responsible for checking that the summary is accurate and that the email reflects the actual conversation; this builds trust on both sides. As you see patterns, you may create specialized prompts for different deal stages or segments, further speeding up the process.

Clean CRM data is essential for forecasting and targeting, but keeping it clean is tedious. Claude can review records for missing or inconsistent fields, stalled deals, and duplicate entries. It can then propose fixes, such as suggested values for missing industries, reminders for deals that have not changed in weeks, or merge candidates for duplicate contacts.

In this module, you will define what "good hygiene" means for your CRM: which fields must always be filled, how long is acceptable for a deal to sit without activity, and so on. Then you will experiment with Claude-generated hygiene reports and decide which suggestions should be auto-applied and which should go through human review. The objective is a CRM that reflects reality without rep burnout.

Support teams handle a high volume of incoming messages. Claude can help by reading each ticket, assigning a category and urgency, and drafting an initial response. When connected to a knowledge base, it can pull in relevant policy or help articles. Agents then review, adjust, and send responses, while also correcting any misclassifications.

This pattern reduces response time and makes it easier to route complex issues to the right people. You will practice designing prompts that ask Claude to return both a structured triage record and a suggested reply. You will also discuss when it is safe to let Claude send directly (for example, very simple FAQ questions) and when a human must always approve.

Over time, support teams accumulate patterns: the same questions and the same good answers. Claude can mine historical tickets and replies to suggest new macros or improve existing ones. You will give the model batches of tickets and ask it to cluster them into themes, then write clear, reusable responses for each theme.

Once you have reviewed and approved these macros, you can use them both in your help desk and as few-shot examples when Claude drafts future replies. This creates a loop where human expertise feeds AI, and AI, in turn, helps capture and spread that expertise to new agents.

Many processes live only in people's heads or in a few screen-share videos. In this unit you will learn how to turn those into written standard operating procedures (SOPs). You will start with a Loom or similar recording of a real process, then feed Claude the transcript and ask it to extract a step-by-step guide, including prerequisites, decisions, and troubleshooting tips.

The resulting SOP will need editing, but it gives you a strong first draft. Once reviewed, it can be stored in your knowledge system and used as a basis for training and automation. As processes change, you can update the videos and regenerate updated SOPs, keeping documentation closer to reality.

Contracts can be long and dense. Claude can help you understand them faster by highlighting key terms, unusual clauses, and obligations in plain language. You will learn to give Claude a contract and ask it to summarize main points, note differences from your usual templates, and list questions you might want to ask a lawyer.

Claude is not a replacement for legal advice. Its role is to make the document more accessible and to surface areas of concern quickly. You will practice reading Claude's summaries alongside the original contract and comparing them to your own understanding. This makes conversations with counsel more efficient and ensures fewer surprises after signing.

When a new hire joins, they need a clear path from "day one" to "fully productive". Claude can help create structured onboarding plans that combine role goals, training materials, shadowing, and practice tasks. You will start by defining what "productive" means at 30, 60, and 90 days for a given role, then list the resources and activities that support those milestones.

Claude will take this information and draft a week-by-week plan, including reading, meetings, exercises, and check-ins. Managers will then adjust the plan based on their knowledge of the person and the team. This makes onboarding more intentional and repeatable across roles and offices.

Employees often have recurring questions about policies, benefits, and procedures. A policy Q&A bot built on Claude can answer many of these quickly by reading your HR handbook, benefits summaries, and internal FAQs. In this unit you will select a set of official documents and ask Claude to answer common questions using only those sources.

You will design prompts that instruct Claude to say "I don't know" or escalate when the answer is unclear or not covered by policy. You will also think about privacy and fairness—making sure the bot does not make up answers about sensitive topics. Done well, this frees HR staff for more complex work while giving employees faster access to reliable information.

Ecommerce teams need many product listings across channels, each with slightly different constraints. In this unit you will use Claude to generate titles, descriptions, feature bullets, and FAQs based on structured product data and your voice doc. You will be explicit about which attributes to highlight and what claims are allowed.

After Claude generates listings, you will check them for accuracy (no invented features), compliance (especially in regulated categories), and performance (click-through and conversion over time). You can then refine your prompts and templates based on which listings perform best.

Customer reviews contain a lot of insight but are hard to digest at scale. Claude can read many reviews at once and cluster them into common themes: what people love, what frustrates them, and what they wish existed. It can also extract representative quotes you can use in marketing.

You will practice feeding Claude raw review text (with personal details removed where appropriate) and asking for a summary organized into strengths, weaknesses, and opportunities. You will then translate those themes into concrete actions for product, support, and marketing. This closes the loop between customer voice and business decisions.

QuickBooks holds sensitive financial data. Your first integrations will therefore be read-only: Claude can look but not touch. You will connect a limited set of reports—such as your P&L, balance sheet, and AR aging—to Claude and design prompts that ask for explanations and narratives, not changes. For example, rather than "fix these transactions", you will say "Explain the main reasons gross margin changed month-over-month".

In this unit you will learn to scope permissions narrowly and to log which reports Claude accesses. Over time, you may choose to let the AI suggest categorizations or draft memos for specific entries, but humans remain responsible for posting actual transactions. This pattern keeps financial systems accurate while still benefiting from Claude's ability to explain and summarize.

With HubSpot, Claude can help you maintain a healthy pipeline by analyzing opportunities, spotting stalled deals, and suggesting next actions. You will first connect Claude in a way that allows it to read deal data but not update it directly. Then you will design prompts that ask for pipeline overviews: which deals are stuck, which may be overvalued, and where to focus attention this week.

Claude can also propose fields to standardize and contacts to merge, but you decide which suggestions to accept. As with QuickBooks, logging and clear ownership ensure that everyone understands which changes are human-driven and which were AI-assisted. This keeps your data trustworthy and your team confident.

Google Workspace contains much of your organization's written memory: docs, slides, emails, and calendars. Claude can help you search, summarize, and act on this information, but only within the boundaries you set. In this unit you will decide which folders and labels are safe for Claude to read and which should remain private.

You will then design workflows such as a daily or weekly briefing that summarizes upcoming meetings, key email threads, and outstanding docs that need attention. Claude will draft this briefing using your calendar and email metadata, and you can refine both the prompts and the selection of sources to match your team's needs and privacy expectations.

If your company uses Microsoft 365, you can apply the same patterns with Outlook, OneDrive, SharePoint, and Teams. Claude can summarize long email threads, collect documents from a project folder, and produce meeting notes and follow-ups. The key principles—least privilege, human review, and logging—remain the same.

You will pick one or two scenarios, such as "preparing for a weekly team meeting" or "reviewing project documentation", and practice designing Claude workflows around them. This helps you bring AI assistance into your everyday tools without disrupting existing permissions and compliance settings.

Slack is a natural place to approve or adjust Claude's suggestions. Instead of acting silently, Claude can post proposed emails, contract summaries, or AR actions into specific channels. Humans then respond with simple signals—such as emoji reactions or short comments—to approve, edit, or reject each suggestion.

In this unit you will define which workflows must go through Slack, which channels will host approvals, and who has authority to sign off. You will also think about how long items can sit before escalation. This turns Slack into an "approval queue" that keeps humans firmly in the loop while still benefiting from automation.

DocuSign holds your agreements. Claude can read documents sent through DocuSign and highlight unusual clauses, missing standard terms, or obligations you should be aware of. You will learn to ask for structured summaries that list payment terms, termination conditions, liability caps, and other key points, with page references so humans can verify them.

You will keep Claude in a read-only role: it does not sign or send anything. Legal or leadership remains fully responsible for decisions. Claude's value is in making reviews faster and more consistent, especially for non-lawyers who need to understand what they are signing.

Notion and similar tools act as a "second brain" for many teams. Claude can help keep that brain organized by creating, updating, and summarizing pages and databases. For example, you might ask Claude to read a project's meeting notes and update a status page, or to turn a collection of research notes into a structured overview.

You will define which spaces Claude can read and write to and how you want it to structure content. You will also establish a habit of reviewing AI-generated pages and cleaning up as needed. The goal is a knowledge base that grows and stays useful without requiring constant manual upkeep.

All integrations rely on sound security and governance. In this final integrations module, you will focus on permission scoping and audit logs. Least privilege means Claude only gets access to what it needs for a given workflow—no more, no less. Audit logs record which files and records the AI has accessed and what actions it has taken or proposed.

Together, these practices allow you to trace issues back to their source and show regulators, partners, or internal stakeholders that you are in control. You will sketch system diagrams that show where Claude can see and act, and define a simple routine for reviewing logs, such as a monthly or quarterly check-in with someone from operations or security.

In this course, "agentic" means more than a single prompt and answer. An agentic workflow is a system in which Claude can take multiple steps, call tools, and make decisions within rules you define. It might read data from one system, transform it, decide which items need attention, and then draft actions in another system, all in one run.

Agentic workflows are powerful, but they also introduce new risks: if something goes wrong in step three, the problem can carry through to later steps. That is why you will always combine agentic behavior with checkpoints, permissions, and logging. In this section you will learn to think like a system designer, not just a prompt writer.

A multi-step run breaks a complex task into stages: gather data, analyze, propose actions, and prepare outputs. Checkpoints are moments where the run pauses and waits for a human or another system to approve, adjust, or add information before continuing. In this module you will practice drawing workflows as step-by-step diagrams, marking where each checkpoint sits.

For example, a marketing agent might first compile performance data, then draft a report, then wait for a human to approve that report before drafting follow-up experiments. A finance agent might prepare a cash briefing and then pause for sign-off before sending it to leadership. These checkpoints balance automation with control.

Giving Claude write access—to change records in your CRM, post transactions in accounting, or edit documents—raises the stakes. In this unit you will learn a simple way to decide when, if ever, to grant write access. You will consider how critical a system is, how easy it is to roll back changes, how well you can log activity, and how mature your evals and guardrails are.

Often, the safest path is to start with read-only access and human-mediated writes: Claude proposes changes, and humans apply them. Over time, for low-risk and easily reversible actions, you might allow Claude to apply changes directly within narrow limits. This gradual approach keeps you from jumping too quickly into fully automated actions where the impact of mistakes is high.

No system is perfect, and agentic workflows are no exception. In this module you will list ways an AI-driven workflow could fail: acting on stale data, misinterpreting instructions, sending messages to the wrong people, or partially completing tasks without telling you. For each failure mode, you will design a way to detect it and a way to roll it back.

Rollback patterns include keeping a log of all AI-initiated changes, offering "undo" options, and designing operations so that they are idempotent—running them twice does not cause harm. You will also discuss how to simulate failures before going live, such as running a workflow on test data or in a sandbox to see what happens when things go wrong.

Every time Claude runs, it consumes compute and time. Multi-step workflows with long context windows and many tool calls can become slow or expensive if not designed carefully. In this unit you will learn to set simple budgets: how long a workflow is allowed to take and roughly how much you are willing to spend per run, given the value it creates.

You will look at ways to optimize: trimming unnecessary context, reusing summaries instead of pasting full documents repeatedly, batching similar tasks, and using lighter-weight models for non-critical steps. The goal is to design workflows that feel fast and that make clear economic sense.

A good user interface makes it easy for humans to supervise and guide agentic workflows. Common patterns include showing a proposed change alongside the current state (a diff view), presenting a queue of suggestions with simple approve/edit/reject buttons, and allowing users to give structured feedback such as "too generic" or "missed key detail".

In this module you will design simple UIs—often just as sketches—that show how humans will interact with Claude's suggestions in your tools. You will also decide what feedback you want to capture, such as error reports or satisfaction ratings, so you can improve prompts and workflows over time.

The capstone of the course is to design and, if possible, implement one full multi-step workflow that matters to your business. You will choose a use case that is meaningful but not existential—for example, a weekly executive briefing, an AR review loop, or a campaign production pipeline. You will then map the steps, decide where Claude reads and writes, add checkpoints, define success metrics, and outline how you will monitor and improve the workflow.

By the end of this unit, you will have a written specification for your workflow and a plan to pilot it. This becomes the starting point for a new phase: running AI-driven operations in production, learning from their behavior, and gradually expanding where Claude can help.