This workshop uses a task from your own work to explore business context, task definition, and reviewing AI output. You can use Claude or ChatGPT.

Access to connections, skills, and scheduling varies by account. A browser is enough to begin with text or files you're allowed to share.

I use these tools in my own business and help clients choose useful tasks, work through the details, and put the results into practice.

My approach is to work with you: you know your business, and that knowledge shapes the instructions and the review.

A finished automation isn't required.

Defining a task you care about, identifying the information it needs, and deciding what a good result looks like are useful outcomes. If you can try it during the workshop, review the result and record what needs to change.

An agent can gather information, use connected apps, and carry out several steps within an assignment. The interface may still look like a chat box.

You supply the goal, relevant context, and limits, then review what comes back. Tool access doesn't guarantee correct work.

These are three ways to approach the same business problem. Delegation combines several steps: find the inquiries, check what each person needs, update the customer records, and prepare replies.

Define which records may change and which actions need approval. A useful result may include drafts and questions about missing information.

These examples come from my own work. A newsletter skill reads AI newsletters in my inbox and prepares a digest.

After a call or event, an interview captures what happened and supports record updates, reminders, and follow-up drafts. Workshop submissions become reply drafts. I review the proposed work and approve changes or outgoing messages.

Past participants have built a skill for a recurring email marketing task, started improving weekly client reports, and worked on a dashboard drawing from several sources. One attendee told the room an exercise had saved her twenty hours.

That was her report, not a measured result or an expectation for this workshop. Choose a task that matters to you, even if the first step is small.

These are possible uses to test against your own work. For project notes, ask for missing information and who owns each item.

If an owner isn't recorded, the result should say so. For invoice drafts, compare the output with approved work records and agreed rates before using it.

For the task you want off your plate, describe what starts the task, what information you use, and what should come out.

A small task is fine; a large one can be broken into smaller parts. Workshop submissions may be used in the live demo, so leave out confidential information and say if you want yours kept off screen.

For these exercises, use Claude or ChatGPT. The coding products are included to help distinguish the names. As of October 7, 2026, Claude's unified chat and Cowork experience is rolling out to Pro and Max accounts.

Some accounts still show a selector. Product features and access differ by plan; the same task-definition and review habits apply across tools. See the Claude rollout guide for that change.

This is a quick introduction to your business, not an in-depth review. Give short answers and stop once the context is useful for the task ahead.

Check the summary and correct misunderstandings. Ask where it saved the summary and how to use it in another task. You can keep your own copy in a document or project; saving something doesn't guarantee every future conversation will use it.

Continue in the conversation that has your business context. Explain what starts the task, what you normally do, which information you use, and how you judge the result.

The summary should be specific enough to check before implementation. Dictation is fine. For recurring work, you can later keep background files and instructions together in a project.

Where work runs and where files live are separate questions. From October 6, 2026, new Claude Cowork tasks on Pro and Max run in the cloud. Access to connected local folders still requires the desktop app to be open and online. Closing your laptop only leaves the task able to finish if everything it needs remains accessible. Using a desktop app doesn't mean processing stays on your computer. See Claude's cloud and desktop guide.

My daily AI brief illustrates the distinction. Researching and writing the brief is one part; saving the dated file in my own project folder is another. A result in a cloud conversation doesn't confirm that my local archive or backups contain it. Check the output location before calling the task complete.

Local resources can also include a shared drive or server on your workplace network (LAN). Access depends on the connection, permissions, and network reachability. A cloud tool does not automatically gain access to that network.

This overview connects recent cloud changes with other approaches to ongoing assistance. The product descriptions were checked October 6, 2026: OpenAI dots, Meta Muse, and Grok Bot. My reading of Claude's direction is that ongoing context, cloud work, and schedules support a similar kind of continuing assistance. That's my interpretation, not a promise of identical products.

For any option, check what it remembers, which apps it can reach, where results are saved, and when it asks for approval. Open-source setups can offer more control and can be more private, depending on the model, hosting, and connections you choose. They also require more maintenance. You don't need to switch tools to work on your task.

Training settings and access permissions answer different questions. One controls use of conversations for model improvement; the other controls what information a task can reach. Neither should be inferred from the fact that you're using a desktop app.

The source checks for this deck are dated October 5–7, 2026. Consumer accounts have model-improvement controls; business offerings don't train on business data by default. Check your own plan and current settings: OpenAI data use and Claude data use. Turning off training doesn't mean data stays on your computer. Leave out material you aren't permitted to share; use suitable examples from your own work.

The following slides follow one process: read a submission, draft a useful reply, review it, save the working instructions, and test them on another submission.

The exercise demonstrates how business context and corrections shape the result.

A connection gives the agent access to information or actions in another app, such as email or files. Reading email and sending it are different capabilities, though the available permission controls vary.

Start with what the task needs. A “don't send” instruction is useful, but use permission restrictions too when the tool offers them.

The demo uses workshop submissions as inputs and useful reply drafts as outputs. The reply needs enough business context to suggest a realistic first experiment.

If information is missing, it should ask a question. Keep access limited to the intended submissions, and review every draft before using it.

Supply the submission along with relevant business context and writing preferences. Begin with one item so you can inspect the whole result.

If a connection isn't available, you can paste an appropriate example from your own work. Check whether the agent needs more information before expecting a useful answer.

Check whether the suggested experiment is specific enough to try and whether the draft invents facts or promises.

Correct what needs correcting; a good draft doesn't need an artificial flaw. Save useful corrections as instructions so you can test whether the next result follows them.

Create the skill after working through the task and correcting the result. Check that the saved instructions include the inputs, steps, review criteria, and approval limits. Keep general rules separate from individual customer details. Try a different example, including one with missing information, before running a batch.

A saved file isn't proof that a new task can use the skill. Confirm it's available in your tool, then test it in a fresh task. Templates and source files must also be accessible. If the tool doesn't support skills, keep reusable instructions in a document. For reply batches, check recipient addresses and duplicate drafts. In form emails, the attendee's address may be in the form body rather than the sender field.

Saving instructions and arranging for them to run are separate steps. Schedules and event triggers depend on the tool and connections available. Run the task yourself first, check its results, then consider repetition.

Before unattended runs, check every input dependency, who notices a failure, and how to pause the task. Cloud execution doesn't remove a dependency on your computer. Test a run with the access conditions you intend to rely on. Claude's scheduling guide describes its current setup.

The demo's sequence can be reused with another task: supply context, define the result, try one example, review it, and save the useful instructions.

Connections, skills, and schedules can be added as the task becomes clear.

Choose a piece of work small enough to inspect. State what a good result includes and what requires your approval.

Review for supplied facts, missing information, and usefulness. Then try an incomplete or awkward example to see where the instructions need work. Clear instructions help, but results still need checking.

Continue from the interview and use the template to define the assignment. You don't have to connect an app: an appropriate example from your own work can be enough.

If implementation isn't possible today, identify missing inputs or access and describe what a good result would look like. Working out a task you care about is useful progress.

Supply an example and inspect the result using the same checks as the demo. Correct it, save useful instructions, and try another example if you can.

If you're stuck, narrow the task or identify what information is missing. You can leave with a clearer task and next step even if you haven't implemented it.

Compare the result with what you wanted. Record what worked, what needed correction, and what background knowledge you had to supply.

Include the time spent explaining and reviewing when deciding whether to keep using the process. A failed attempt can still reveal a missing input or an unclear task.

Put corrections in the saved instructions and check that later runs follow them. Polished writing can still contain errors. For recurring work, name someone to notice failures or missing inputs, and know how to pause it.

Emails and documents the agent reads may themselves contain instructions. Treat those as source material, not permission to change your assignment. Judge the process by the whole workload, including setup, corrections, and review.

Ask the tool to explain an unfamiliar step or question you about unclear details. For work with several steps, review a plan before it acts.

Ask it to save corrections and check where they were saved. Voice can make it easier to explain your work; review the captured details, especially names and numbers.

Choose a task you care about and record the next step. That might be gathering an example, clarifying the desired result, or trying another run.

Decide what improvement would justify continued use. You don't need every feature covered in this workshop.

I help small teams assess where AI fits, implement useful workflows, and train the people who use them.

The work includes business context, testing, and review. An intro call is a way to discuss your task and see whether working together makes sense.

Use the feedback form to share what helped and what could be clearer.

The opening slide links to this deck so you can return to the prompts and explanations when you work on your task again.