How to create clear, reusable work prompts
A work prompt is an instruction that lets a model prepare a result with a clear objective, context and limits. Its quality is measured in the work you can review and use. This guide shows how to write, test and maintain reusable instructions without relying on magic formulas or ever-longer requests.
By the end, you can:
- Organise the task, context and output
- Handle missing data without inventing answers
- Version instructions with test cases
Replace a vague request with an observable result
“Write an excellent reply” does not define what the team needs. Identify the recipient, purpose and work status: “prepare a draft to clarify missing details in a quote request”. Add what the person should be able to do next. The objective guides information selection and review.
Use specific verbs: summarise, compare, extract, organise or draft. If a request contains several tasks, state the order and result of each. Avoid assigning a grandiose role as a substitute for context. Saying “you are the best expert” does not provide prices, conditions, sources or approval criteria.
Organise five parts that serve different functions
A practical structure includes the task, approved context, case data, output format and criteria. Anthropic’s prompting guidance emphasises clarity, examples and separation of parts. Use simple headings or delimiters the team can maintain. Organisation helps interpretation; it does not work as a security barrier.
Provide the context that changes the answer. An entire history can bury relevant information and increase cost and review. If a task depends on unavailable information, specify how to proceed: flag the gap, request clarification or prepare only the supported part. Define this before demanding a complete answer at any cost.
| Part | Question answered | Example |
|---|---|---|
| Task | What should be prepared? | Draft reply to a rescheduling request |
| Context | Which rules apply? | New time only after checking the calendar |
| Data | What do we know about this case? | Current appointment and stated preference |
| Output | How should it be delivered? | Short message and internal list of pending points |
| Criteria | How should it be approved? | No invented availability |
Try a fictitious rescheduling request
In the fictitious case below, the calendar has not been checked. The objective is to prepare a useful reply without turning a preference into a booking. The instruction separates the customer message from team notes, preventing internal comments from being sent accidentally. The draft remains subject to review before sending.
Task
Prepare a draft reply to a rescheduling request.
Approved context
The new time is confirmed only after checking the calendar.
In this exercise you have no calendar access and cannot send messages.
Fictitious case data
Customer: Ana Exemplo.
Current appointment: 25 November 2026 at 14:30.
Request received: “Can I move to the afternoon of the 26th?”
Result
1. Customer message in British English, up to 80 words.
2. Separate internal notes: details to confirm and next step.
Criteria
Acknowledge the stated preference and explain that it will be checked.
Do not assign a new time or declare the appointment changed.
Do not add fees, deadlines or conditions absent from the context.Check the content, not only the tone
An acceptable draft for the exercise would be: “Hello, Ana. We have received your request to move the appointment to the afternoon of 26 November. We will check availability and confirm an alternative with you. The change remains unconfirmed.” The internal note identifies checking the calendar as the next step. This is an editorial example, not a reply that was sent.
During review, look especially for changes in meaning: promising a time, confirming a change or inventing a condition. Then review clarity and tone. A friendlier message does not compensate for an incorrect commitment. If the case needs system access, that access belongs to the application and permissions, not to the strength of the prompt’s wording.
Use examples that clarify boundaries
Add examples when an abstract rule is insufficient. Show a complete case, a case with missing information and an out-of-scope case, each with its expected result. Identify them as examples and keep them separate from current data to reduce the risk of copying names, dates or values from an old case.
Test the instruction with an ambiguous date, two requests in one message and text contradicting the approved rule. Change one part at a time and keep the result. Asking for more detail, adding context and changing the format simultaneously makes it harder to understand what helped. A model change also justifies repeating the cases.
- Examples should cover different decisions, not merely change the customer’s name.
- A missing value should remain visible as a pending point.
- Output for software requires structure and content validation.
- Instructions should fit a process the team can explain.
Correct the cause instead of accumulating prohibitions
If the answer keeps failing, investigate whether information exists, is current and whether the task suits the model. A larger prompt does not fix an integration that does not check the calendar or a source contradicting another. Remove old instructions that no longer apply and avoid rules simultaneously demanding extreme brevity and exhaustive explanation.
Do not place passwords, keys or unnecessary data in shared instruction templates. Request references or a brief justification a person can check rather than treating a lengthy model explanation as evidence. Verification should return to facts and observable actions.
Maintain a small catalogue of approved instructions
Give regularly used instructions a name, version and person responsible. Keep test cases and explain when to use each instruction. A useful catalogue can contain a few well-maintained instructions; dozens of almost identical variants make it harder to identify the current one. When a business rule changes, update context and tests together.
Worksheet for your project
Task name and version: Person responsible and intended users: Expected result and next action: Approved context and origin: Variable case fields: Output format: Behaviour when information is missing: Included examples and the reason for each: Review criteria: Test cases and results: Application capabilities required: Approval date and change requiring another review:
Sources and criteria
Content prepared with AI assistance and research using technical sources, with checks of the relevant examples and calculations. The references support the guide’s decisions; examples identified as fictitious let you try out the method.
