RUDI/Insights & resources/Prompting guide

A practical guide for workplace AI

Give AI the context
to do useful work.

Describe the task, provide the right material, and decide how you’ll check the result. When AI can use connected tools, define what it may do and where a person takes over.

By RUDI · Updated September 22, 2026

Start with the work you need done.

A project manager asks AI to prepare a weekly update. A useful result depends on the current project notes, the audience, the decisions that need attention, and which details may be shared. A fluent summary of an old document could still be the wrong update.

Prompting now often sits inside a larger workflow: choosing files, using search, connecting an approved account, or asking an agent to prepare an artifact. Your instructions describe the job. The surrounding setup determines what information and actions are available.

01Brief the task
02Add context
03Review the work
04Approve its use

You can begin in plain language. Add detail when it resolves a real ambiguity, and expect to refine the work through conversation. Microsoft’s Copilot guidance describes a similar starting point: a goal, context, expectations, and sources.

Write a brief

Use CRAFT to explain the assignment.

RUDI’s CRAFT checklist helps you make five decisions. Use the parts that matter for the task; a short request can be enough when the context is already clear.

Context: What is happening, who needs the result, and which material should the AI use? Identify the current files or approved sources.
Role: What perspective would help? “Review this as an operations coordinator” gives a useful lens. You still need a qualified person to check work that requires expertise.
Action: What should it produce or change? Name the deliverable and the boundary: draft an update, compare two proposals, or extract dates without editing the source.
Format: How will someone use the result? Specify a short email, a table with named columns, an annotated document, or another practical output.
Tone: Who is reading it? Describe the language they need, or provide an approved example when style matters.

Then add the review criteria: what must be accurate, what to do with missing information, and who approves the result. Those decisions become especially important when the same task will run repeatedly.

A worked example

Prepare a weekly project update.

Instead of leaving the audience and source material implicit, give the AI a brief it can work from. This example assumes you have attached notes that your organization permits you to use in the approved AI workspace.

Draft this week’s project update for our department head.

Use only the attached project notes dated September 21.
Focus on completed work, blockers, and decisions needed.

Return:
• A summary of up to 150 words.
• An action table: action, owner, due date, source reference.
• A separate list of missing or conflicting information.

Keep the language direct. Do not infer an owner or deadline
that the notes do not identify. Mark those fields “To confirm.”

If a required attachment is missing or unreadable, tell me
before drafting. Prepare the update for my review; do not send it.

Review it: open the notes and check each action, owner, date, and source reference. Confirm that the summary includes the decisions the department head actually needs to make.

Then refine it: “The vendor decision needs to be more prominent. Move it to the first paragraph, keep the original deadline, and leave the action table unchanged.” A specific correction is easier to assess than asking for a “better” version.

Supply context

Choose the material the task depends on.

Start with the approved source of truth: the current policy, relevant worksheet, project folder, or email thread. Label versions and dates. If documents disagree, identify which one governs or ask the AI to surface the conflict for a person to resolve.

Use an approved connector when the task needs information from another system. Confirm which account is connected, what it can access, and whether the relevant file was actually retrieved. A link in a prompt does not establish that the AI can open it.

Show an example when it helps explain an unfamiliar format. Remove information the task does not need, particularly personal or confidential material. Keep the task instructions visibly separate from the documents being analyzed.

OpenAI documents how relevant context and clear structure support a response; Google’s prompt design guidance shows how examples can establish a desired format. Neither replaces checking the output against the source.

Try a work task

Adapt a brief to your own workflow.

Replace the bracketed fields with your task details. Use approved or fictional material for practice.

Compare proposals for a decision

Compare [proposal A] and [proposal B] for [decision].
Use these criteria: [criteria agreed by the team].

Make a table with each criterion, evidence from each proposal,
and questions we need answered. Include page or section references.
Keep quoted prices, units, and assumptions distinct.

Finish with the trade-offs and information still needed.
Do not choose a winner where a required fact is missing.

Draft a response from an approved policy

Draft a reply to [request] using [approved policy and version].
The reader needs to know [what they can do next].

Keep it under [length] and use [tone]. Identify the policy section
supporting the reply. Flag any part the policy does not address.
Keep personal information out unless it is needed and approved.

Return a draft for [reviewer]. Do not send or update any record.

Map a process before building an agent

Use these process notes to describe how [task] works today.
List the inputs, steps, systems, decisions, handoffs, and output.
Mark anything not established in the notes as a question.

Suggest where AI could help prepare work for a person to review.
For each suggestion, identify the data and permissions required,
likely failure cases, and who would be responsible for the result.

Do not assume a connector exists or change any system.

Connected tools & digital workers

Define what the AI may do.

Once AI can send a message, update a record, or run code, the assignment needs an action boundary. Identify the permitted systems, the changes it may prepare, the actions that need approval, and the conditions that require it to stop and ask for help.

Allowed work: For an email workflow, read the approved folder and prepare a draft using the current response policy.
Human approval: The assigned reviewer checks the recipient, content, and attachments before anything is sent.
Exceptions: A missing policy, conflicting record, unusual request, or inaccessible file goes to the human owner. The workflow records the incomplete step.
Actual controls: Restrict account permissions and available tools, configure approval gates, and keep an activity record. Test that these controls work.

A sentence in a prompt cannot enforce a permission boundary. Emails and documents can also contain instructions that try to redirect an agent. Configure access and approval controls in the system itself; see OpenAI’s guidance on prompt injection and tool approvals.

RUDI’s RESPECT framework starts with a responsible human. Every digital worker needs someone who owns the workflow, reviews problems, and can pause it. Learn about managed digital workers and AI workspace management.

Review & improve

Check the result against the job.

Check the evidence. Open cited sources. Verify names, numbers, dates, calculations, and any claim someone will rely on.
Check completeness. Look for missing tasks, exceptions, and unanswered questions. A polished response can still omit a requirement.
Give a specific correction. Name the error, provide the missing context, or show the expected output. Recheck affected parts after the revision.
Test before reusing. Try a typical input, one with missing information, and a difficult case. Keep a record of what failed and what you changed.
Keep an owner. Save useful instructions with their sources, review checklist, and revision date. Revisit them when the policy, tool, or workflow changes.

Anthropic recommends defining success criteria and ways to test them before optimizing a prompt. For a team, this can begin with a few representative tasks and a reviewer who knows the work. If the result takes more effort to correct than the task used to take, reconsider the workflow.

Adapt the workflow to your approved tool.

The same work brief can be a useful starting point in Microsoft Copilot, Gemini, ChatGPT, or Claude. File access, connected tools, saved instructions, and approval controls vary by product, plan, and configuration. Check what your workspace actually supports.

For coding and application work in tools such as Codex or Claude Code, add the relevant project instructions, the files the agent may change, and the checks the work must pass. Have a qualified reviewer inspect the change before release.

Use the vendor’s current documentation when you need a particular feature. Test changes with your own tasks rather than relying on a universal list of model tricks.

Common questions

Do I still need to learn prompting?

Yes. Describing a task, choosing relevant context, and checking a result remain useful skills. You can practice them through ordinary conversation with an approved AI tool.

Should I ask AI to explain its thinking?

Ask for the evidence, assumptions, calculations, or a brief explanation you can check. A plausible explanation is not proof that the answer is correct.

Can a prompt prevent factual errors?

No. Clear sources and instructions help, but the response still needs review. Verify important information against the original material and ask a qualified person to review work that requires specialist judgment.

When should a saved prompt become a digital worker?

Consider it when a recurring task has clear inputs, a defined output, permitted tool access, and a person responsible for exceptions. Test the full workflow, including failures and approval steps, before scheduling it.

Practice with your team

Bring a real task
to the training.

RUDI helps people use AI in their work, review what it produces, and identify where an agent or small application could help. Start with the tools your organization approves and the tasks your team actually performs.

Explore team training Work through a department workflow