PRACTICAL AI AT WORK · ANTHROPIC · CLAUDE
The most useful first AI assistant is often quite small. It might organise an enquiry, turn approved notes into a briefing or help someone find the right section of a procedure. A narrow job is easier to explain, test and maintain than an assistant expected to “help with everything”.
Claude’s project and artifact capabilities offer ways to organise ongoing work and create reusable outputs, including documents and small interactive tools. The exact experience depends on the account. Think first about the work you want to make repeatable, then choose the appropriate feature.
Define the assistant’s job in one sentence
For an illustrative training enquiry assistant: “Convert an approved enquiry into a structured needs summary and a list of questions for a training adviser.” That is a useful boundary. It does not include quoting a price, promising a date or deciding that a customer is eligible for funding.
Write down the expected inputs, output structure and escalation conditions. Include what should happen when the customer provides too little information or asks for something outside the service.
Provide examples of good work
Give the assistant a few anonymised sample enquiries and reviewed summaries. Add current reference information and explicit instructions about uncertainty. Examples often reveal details that a general instruction misses, such as how your team distinguishes an individual booking from a company-wide requirement.
Keep the references manageable and current. More material is not automatically better. Conflicting versions can produce a confident mixture of old and new information.
A PRACTICAL WORKFLOW
A prompt for the first prototype
Help me design an enquiry assistant. Its only job is to produce a structured summary for a human adviser. Extract the stated topic, audience, number of learners, delivery preference and timescale. Mark missing fields as unknown. Suggest up to three clarification questions. Never invent a price, date or commitment. First show the proposed output format and five test cases, including incomplete and out-of-scope enquiries.
Use an artifact when an interface helps
A small form or interactive checklist may be easier for colleagues to use than a long conversation. Claude can help create that prototype as an artifact. Label it clearly as a prototype and test it with sample data before introducing real customer information.
A convincing screen is only one part of a working system. Persistent storage, access control, integrations, logging and support need separate attention. Do not assume that a generated interface has safely implemented those requirements because it looks finished.
Test the awkward cases
Try a normal enquiry, an incomplete one, a contradictory one and a request outside the assistant’s role. Include text that tells the assistant to ignore its instructions. The expected response should preserve the defined task and flag the problem, rather than treating customer-supplied wording as permission to change its behaviour.
Record the expected result for each test and compare it with what happened. Decide who can approve a release and who will handle a failure. Retest after changing the instructions or source documents.
Plan for ownership
Assign someone to maintain the references and review usage. Start with a person checking every output. Expand autonomy only when the workflow, permissions and evidence justify it.
ChatGPT and Gemini also provide ways to organise reusable instructions and work. The transferable skills are defining the job, designing the input and output, testing exceptions and understanding what still needs engineering. Those are practical skills a business team can learn without pretending that a prototype is a finished production system.
Sources and further reading
Product information checked 4 November 2026. Features and availability depend on your plan and administrator settings. Examples and suggested workflows are illustrative.