Practical AI for Executives

Navigate to Settings > Capabilities

In the Memory section make sure that "Search and reference chats" and "Generate memory from chat history" are both toggled "On" (the button turns blue).

You should see an option to "View and manage memory" appear if you need to toggle "On" the above options.

Start a new chat and enter a prompt that establishes standing context.

Verify the new entry appears.

Edit or remove anything that is incorrect, then reflect:

Return to a new standalone (non-incognito) chat.

Ask Claude to retrieve content from the memory chat you ran earlier in this lab.

Watch the chat for a visible tool call indicating Claude is searching your chat history.

The result should include a citation linking back to the source conversation.

Set the model to Claude Haiku 4.5. Submit the prompt. Note:

Start a new chat. Set the model to Claude Sonnet 4.6. Submit the same prompt. Compare to Haiku:

Start a new chat. Set the model to Claude Opus 4.7. Submit the same prompt. Compare to both prior runs:

Which model produced the most usable response for this specific task?

Was the difference in depth proportional to the difference in latency?

For your daily work, which model would you default to? When would you escalate to Opus 4.7? When would you drop to Haiku 4.5?

Begin by submitting a deliberately weak prompt and observing the response.

Pick one concept (actions OR information), strip the filler, and add the missing specifics.

Compare the two responses.

Reflect:

Compare the new output to your C-L-E version.

Reflect:

Upload market_entry_strategy.txt using the "+" button or by dragging the file into the chat input.

Confirm the file is attached

Verify each claim in the model’s response against the source document to ensure Claude only uses information from the document.

Then ask a follow-up question that the document does NOT actually answer.

Review the response to see if Claude correctly identifies that this information is not in the document, or if it hallucinates an answer based on its training data.

Reflect:

Were the citations clickable and to credible sources?

Did Claude cross-reference facts across multiple sources, or rely on a single one?

When would web search be the wrong tool for the job?

Expand the thinking section above the answer to see Claude’s exposed reasoning.

Reflect:

Enable Research mode via the "+" button

The magnifying glass icon turns blue when on

Expand the chain-of-thought log and review it.

Reflect:

Start a new chat and submit a casual, open-ended request that depends on knowing who you are.

Submit and review.

Click your initials in the lower-left corner

Open Settings and choose General. Your profile settings are located at the top of the page.

If your profile is blank, answer the questions presented to you and include instructions to Claude on how you like to receive information, what your role is, etc.

Or edit the existing text if you already had profile settings

In either case:

Resubmit the same baseline prompt.

Reflect:

Pick a single business prompt to use as your control.

Once each under Concise, Formal, Learning, and Explanatory.

Switch styles via the "+" button in the lower-left corner of the chat input area, then Use style.

Which style produced the shortest response? Which produced the longest?

Which style would you send to a busy executive? Which to a new team member who needs to learn the concept?

Did any style misfit the prompt — too formal, too breezy, too pedagogical?

Can you detect a difference between Explanatory and Learning, which are similar but not identical?

In the chat interface, click "+" (lower left of the chat input area)

Then select Use style

In the style dropdown, click Create & edit styles to open the Customize your styles modal

Click Create custom style in the bottom left corner of the modal

Choose the description method

Click Describe style instead link to switch from the upload method to the describe method

Select Tailor to an audience

In the text field on the right, enter one of the following audience descriptions, depending on the style you want to build:

Click Generate style

Review the generated style instructions in the preview panel

Rename the style (e.g., Financial Brief or Data Center Energy Brief) and save

Open a new chat

Return to "+", select Use style, then select your newly created style to apply it to the conversation

Submit the test prompt that matches your audience choice:

How closely did Claude’s generated style instructions match what you would have written manually?

Did the response sound like it was written for the persona described in your audience description?

What would you tweak in the audience description to sharpen the voice further?

Did the email lead with the bottom line, as your audience description specified?

Were technical terms or acronyms used appropriately and defined where needed?

Would you actually send this email, or does it need a revision?

In the left sidebar, click Projects, then click + New Project.

Name the project something descriptive, for example, Pulse Launch — May 5 GA.

Add a brief project description (this is for your organization only; Claude does not see project names or descriptions).

Inside the project, locate the project instructions area and click to edit.

Paste in the following instructions, or adapt them to your own context.

Locate the Files section in the right side contextual panel (this right side panel now contains project details).

Upload the three provided files: pulse_launch_plan.txt, pulse_launch_raci.txt, and pulse_launch_readiness_criteria.txt.

Confirm all three files appear in the file area before proceeding; you should see thumbnails and filenames in the Files area of the right contextual panel.

Did Claude reference specific items from the uploaded readiness criteria, or did it generate generic answers?

Did the response follow your instructions on tone, structure, and naming accountable owners?

If anything looked wrong or invented, which document would have given Claude the correct information?

Clicking New Chat within the project (not at the top level).

Ask a different question that builds on the launch context.

Did Claude maintain the Program Manager persona and tone established in your project instructions?

Could you tell that Claude was drawing from the same knowledge base as in your first conversation?

What advantage does this give you over starting from scratch in a standalone chat?

In the left sidebar, click Artifacts under the main navigation.

You will see a list of any artifacts you have previously created (you likely have none).

Click New Artifact.

A category picker will open.

Choose a category that resembles a tracker, checklist, or dashboard if one is available; or click Build from scratch if no category clearly fits.

The artifact creation conversation will open.

Upload pulse_launch_readiness_criteria.txt to this conversation using the "+" button or by dragging the file into the chat.

Here’s an example artifact prompt to enter into the conversation:

Submit the prompt.

Once it is generated, return to the left sidebar.

You should see the new artifact listed under Artifacts with the name you gave it (or a default name if you didn’t specify one).

Does the artifact contain all 21 criteria, organized correctly across the five categories from the source document?

Does the running count and the Go / No-Go indicator update correctly when you change a criterion’s status?

Is it usable as-is, or does it need refinement?

Here’s an example refinement prompt:

Submit the prompt

Watch Claude generate a new version of the artifact rather than replacing the original.

At the top of the artifact panel, locate the version selector.

Navigate back to the original artifact and forward to the new version

Compare the two versions to confirm Claude built on the original rather than rebuilding from scratch

Revert to the original version if you prefer it, then move forward again to the latest

How does Claude indicate which version is currently active in the panel?

When in your real work would you want to keep an earlier version of an artifact rather than overwrite it?

If you wanted a third version that combined ideas from both, how would you prompt Claude to create it?

Think about mistakes that are operationally, reputationally, or legally consequential.

Capture your thoughts on paper or in a document.

For each of the three subtasks, write a one-line rationale for why a failure there would be particularly consequential.

Open Claude Chat in a new conversation

Stress-test example prompt:

Capture the top three to five risks.

Compare them to the subtasks you identified.

Reflect:

Manually (i.e. using your human brain) assign a delegation category to each subtask.

The four delegation categories:

Use the risks you surfaced in the previous activity to inform every classification.

Use this structure:

| Subtask                          | Category              | One-line rationale    |
|----------------------------------|-----------------------|-----------------------|
| 1. Refine job description         |                       |                       |
| 2. Resume screening              |                       |                       |
| 3. Candidate outreach drafts     |                       |                       |
| 4. Phone screen scheduling       |                       |                       |
| 5. Phone screen question prep    |                       |                       |
| 6. Conducting phone screens      |                       |                       |
| 7. Synthesizing panel feedback   |                       |                       |
| 8. Salary benchmarking           |                       |                       |
| 9. Final hire/no-hire decision   |                       |                       |
| 10. Drafting rejection emails    |                       |                       |

Work through all 10 subtasks. For each, pick a category and revise your one-line rationale if needed, so that it ties back to a specific risk or to the level of human judgment the task requires.

Once you have finished, your completed table will likely have:

Identify your two or three most uncertain classifications — the ones where a category was genuinely hard to pick.

Stress-test each one with Claude:

Create one prompt that includes the entire table of subtasks, your human-assigned delegation category, your rationale, as well as the category descriptions

In that single prompt, ask Claude to stress-test the entire table

Adjust your classifications based on Claude’s pushback, or hold your position with a sharper rationale.

Reflect:

Looking at your final mix:

Step: the subtask

Owner: AI, Human, or Both

Output: what gets produced

Checkpoint: who reviews the output before the next step, and what they look for

Fill in the four fields for each.

For Human only subtasks, you can write a single line summarizing the step and skip the Checkpoint field.

How many human checkpoints did you end up with?

If you had to remove one checkpoint to save time, which would you remove? Which would you keep no matter what?

Could you defend this workflow to your General Counsel?

Read it side by side with the AI-generated brief.

Look for any metric, quote, or external citation in the brief that you cannot match word-for-word against the source

The claim — exactly what the brief says

Why it is wrong — what the source actually says (or does not say)

The failure mode category — choose from the list below

Upload both the actual cascade earnings report and the AI generated brief to the chat

Submit a verification prompt

Reflect:

Explicitly restricts Claude to the uploaded document

Tells Claude what to do when information is not in the document (rather than guessing or inferring)

Forbids external citations or references

Specifies an output format that makes unsupported claims harder to slip in

Open Claude Chat in a new conversation and upload actual_cascade_q3_earnings.txt.

Submit your revised prompt

Compare the new brief against the original flawed brief

Reflect and look specifically for:

In Claude Desktop, click Cowork in the sidebar.

Click the dropdown below the prompt area and select Projects > Create new project.

Click Start from scratch.

Give the Project a clear, descriptive name that reflects its purpose.

Leave the instructions blank for now.

Choose a folder to be your project location.

Click Create to set up the project workspace.

On the project page, click the pencil icon inside the Instructions section.

Write instructions that define how Cowork should handle every task in this project. Tailor them to your chosen workflow. For example:

Save the custom instructions.

Prepare a reference document that Cowork should use as context for tasks in this project. Examples:

Upload the file to the Project, either by putting it in the designated folder or by dragging and dropping it into the Project interface.

Verify the file appears in the Project’s resources.

Review the Project’s memory and context settings.

Add any persistent context that should carry forward across tasks. Examples:

If you have a connector configured in Claude Desktop, tasks within this Project can pull live data from the connected service.

Within the Project, submit a task that exercises the configuration you just set up. For example:

When the task completes, review the output and verify:

If the output does not reflect your project configuration, revise the custom instructions and run the task again.

Within the same Project, define a second related workflow. For example:

Write a task prompt for the second workflow and submit it.

Compare the output of both workflows:

Note any changes you would make to the project instructions to better serve both workflows.

Draft a template block that includes:

Add a short Adaptation Guide with instructions for how another team would customize the template for a different domain.

Test your template by running a third task prompt that differs from the first two workflows. Verify that output quality remains consistent.

Record two template limitations and one improvement you would make before rolling it out to a larger team.

Decide on a recurring deliverable you want Cowork to produce automatically. Choose one of the following or adapt one to your situation:

Write down what the output should include. Be specific. For example, a daily morning briefing might include:

Select Cowork in the left sidebar of Claude Desktop.

Click the Scheduled menu item on the left sidebar.

Click the New task button in the upper-right corner.

Give the task a clear, descriptive name that reflects its purpose and cadence. This is the name the task will saved as. It should be short and memorable. Note that spaces in this name will be converted to dashes and the name will be saved with all underscore letters, so you may want to pick a name that fits this convention. For example:

Write a brief description for the task. For example: "Generate a daily briefing every morning."

Write a prompt that describes the output requirements you defined in the previous step. Be as specific as possible about what the deliverable should include and how it should be structured. For example:

Set the frequency for the task to match your desired cadence. For example:

Click Save to save the scheduled task. ==== Run and Review

If the next scheduled run is not imminent, trigger the task manually to see the output immediately.

Review the output carefully:

Document your observations. Note at least one thing that worked well and one thing you would change.

If needed, update the prompt to address any gaps you identified.

Trigger the task again to verify the improvement.

Define measurable acceptance criteria for your scheduled task’s output. For example:

Add the acceptance criteria directly to the task prompt so Cowork knows what standards to meet.

Run the task again and evaluate the output against your criteria:

Refine the prompt to address any gaps and run it once more to confirm the improvement.

Add an escalation clause to the prompt that tells Cowork what to do when required data is unavailable or incomplete.

Define three failure conditions and the expected behavior for each. Example conditions include missing calendar events, inaccessible connector data, or conflicting source information.

Run the task once with normal conditions and once with a simulated failure condition (for example, by restricting available context). Compare outputs.

Document what changed in the output quality and whether your failure handling instructions reduced risk.

Decide what type of task you want to standardize. Choose one of the following tasks and descriptions, or adapt one to your work:

Write the skill’s instructions. These should define exactly what Cowork must include in the output every time. For example, a Meeting Notes skill might specify:

In Claude Desktop, click Customize in the sidebar.

Click Skills in the left sidebar.

Click the "+" button and select Create New Skill.

Create a new skill with the name (which must be lowercase and have no spaces), description, and instructions you defined.

Save the skill by clicking the Create button.

Confirm the skill appears in your Skills list on the Customize page.

Prepare a sample input for your skill. For a Meeting Notes skill, write or paste a set of rough meeting notes — informal, unstructured, the kind you would jot down during a real call.

Submit a task that uses the skill. For example:

Verify the output follows every instruction in the skill definition:

Prepare a second, different sample input for the same skill. Use different content — a different meeting, a different email, or a different document.

Submit the task again with the new input and the same skill.

Compare the two outputs side by side:

Write a generic prompt for the same task without referencing the skill — for example, "Summarize these meeting notes".

Submit the task using one of the previous inputs.

Compare the unskilled output to the skilled output:

Add a new requirement to your skill. For example:

Save the updated skill.

Run the task again with one of the previous inputs and verify the new requirement appears in the output.

Consider how this skill could be shared across a team:

Define a naming convention for skill versions (for example, meeting-notes-v1, meeting-notes-v2) and when a new version should be created.

Identify three change types and classify them as minor, major, or urgent.

Write a short rollout plan that includes:

Create one new version of your skill and run the same input through both versions. Note the output differences and which version you would recommend for default use.