A practical Claude starter workflow for teams

A useful first exercise in Claude is a task the participant already understands. Rewriting rough notes into an internal update lets people compare the output with the source and learn where an assistant needs direction.

The exercise below can also be adapted to another approved assistant. Product features and account permissions vary, so it relies on supplied text rather than assuming access to connectors, shared projects, or web research.

The provider guidance linked here was checked on September 9, 2026. It comes from Anthropic's developer documentation; the exercise uses ordinary pasted text and does not require an API integration. Model-specific settings and chat-app features should be checked separately for the account you use.

Choose a bounded input

Use fictional or approved notes about a routine project update. Include a completed action, an unresolved question, and a proposed date that has not been agreed. Provide the intended audience and explain where the update will be used.

Keep restricted information out of the exercise. Confirm your team's tool and data policy before moving from practice material to real work.

Give the assistant a clear task

Anthropic recommends explicit instructions about the desired output and constraints. Its prompting guide also describes XML tags as a way to distinguish instructions from source material. The tags below serve that organizational purpose; they are not a security boundary or a guarantee of accuracy. Anthropic's prompting best practices.

Paste this complete example into an approved Claude conversation. The notes are fictional, and the task ends with a draft for review.

<task>
Prepare a short internal update for the website delivery team.
Use only the notes below. Preserve uncertainty and dependencies.
Do not turn a proposal into an agreed deadline.
</task>

<notes>
- Priya approved the homepage headings on Tuesday.
- Might send the draft Friday if the photographs arrive.
- The team has not decided who supplies the photographs.
- Sam asked whether the contact form needs a phone field.
</notes>

<output>
Use three headings: Completed, Open questions, Proposed next steps.
For each item, include the note that supports it.
If an owner or date is missing, say it is unconfirmed.
Do not invent decisions or assign work. Keep it under 180 words.
</output>

Read the result before asking for a more polished version. Styling an inaccurate draft can make the underlying error harder to notice.

Work through an illustrative correction

Imagine a note says, “Might send the draft Friday if the photographs arrive.” If the output says, “The draft will be delivered Friday,” correct the change in meaning.

Ask the assistant to preserve both conditions: the date is tentative and the photographs are a dependency. Then inspect the revision. The learning objective is recognizing that the original sentence contained more than a date.

An acceptable rendering of that item would be: “Proposed: send the draft Friday if the photographs arrive. Timing is tentative; the photograph supplier is unconfirmed.” This is an editorial example of the intended meaning, not a recorded Claude response.

Check the other items just as closely:

NIST identifies confidently incorrect generated content as a risk in its generative AI profile. That is why the checking step remains part of the task. NIST Generative AI Profile.

Save the method with an example

Keep the instructions, an acceptable output, and a short checklist together. The checklist should cover factual support, uncertainty, missing information, and suitability for the audience.

Practice with a different note set before deciding the method is reusable. Record corrections and total effort, including review time.

Anthropic's prompt engineering overview starts with success criteria and a way to test them. For this exercise, success means a reviewer can trace every item to the notes and find no invented commitment. A pleasant writing style is an additional check, not the acceptance standard by itself.

Keep the model name shown in your account, the date, and the prompt with your practice record. If you change models or revise the instructions, rerun both the straightforward and ambiguous examples. Consult the Claude Platform release notes for API changes; confirm chat-app availability in the product's own help before planning a team exercise around a new feature.

For a facilitated version, see the illustrative workshop agenda. AI Training for Teams can be scoped around the work your team actually performs.