Start with the task, evidence and acceptance test
OpenAI advises clear, specific prompts with enough context, followed by review and refinement.1 Turn that into a practical writing habit: state the business task, give the approved evidence and explain what a satisfactory answer looks like. A role such as 'act as a marketing expert' can shape the voice, but it does not supply missing product facts or establish professional competence.
Before prompting, identify the decision the output supports. A supplier comparison needs differences and unanswered questions. A guest email needs accurate conditions and a suitable tone. A board summary needs traceable facts. Once the purpose is clear, ask for a format that makes errors visible. A short table often exposes missing evidence more readily than a polished narrative.
Use this reusable brief in your own words: 'Prepare a draft for this reader and purpose. Use the attached approved source. Return the requested headings or table. Keep supplied facts unchanged. Separate assumptions and missing information from the draft. Ask me about a gap that prevents a reliable answer.' This structure is a working method, not a vendor guarantee.
Specify New Zealand English and the context that matters: the local audience, currency labels, time zone, business terms and the source version. Avoid asking ChatGPT to invent the context from a place name. A proposal for a Wellington customer needs the actual scope and evaluation criteria, not merely the instruction 'make it sound local'.
- Task: name the action and the business purpose in the opening instruction.
- Evidence: attach the approved material or identify the precise connected source.
- Result: state the structure, audience and tone that the reviewer needs.
- Check: request a separate list of unsupported points and assumptions.
Worked customer reply: preserve the terms
Consider a generic Nelson lodge with a public cancellation policy. A guest asks whether a changed travel plan qualifies for a refund. Use a fictional enquiry for practice and an approved copy of the policy. The prompt should draft an explanation, while leaving the booking-specific decision to the person who has the reservation record.
Prompt: 'Using the attached current cancellation policy, draft a plain-English reply to this fictional guest enquiry. Explain which policy clause could apply. Do not decide whether the guest is entitled to a refund because the booking facts are not supplied. Preserve all conditions from the policy. Put the draft first, followed by Missing booking facts and Policy evidence. Use New Zealand spelling and a calm, respectful tone.'
A useful result might say: 'We need to confirm the booking conditions and the date of your cancellation before we can explain the available options.' The evidence section should point to the applicable clause. If the draft says the guest will receive a refund or changes the cancellation window, reject it. Compare every condition with the policy before a manager sends the message.
Refine a weak answer with a targeted follow-up: 'Your reply promises a refund that the source does not establish. Remove that promise. Show which supplied clause supports the revised explanation and identify the facts still needed.' This names the defect and asks for a repair. Asking only for a more professional tone can leave the wrong promise intact.
- Confirm the attached policy is current before generating a reply.
- Check that the draft distinguishes a policy condition from a verified booking fact.
- Keep the real guest's identifiers and private circumstances out of the practice exercise.
- Have the authorised staff member make and communicate the final booking decision.
Worked spreadsheet prompt: specify the calculation
For an illustrative Hamilton retailer, prepare an approved category-sales file with no customer identifiers. Use descriptive columns and a single record per row. OpenAI recommends clear headers and structured data for analysis; ChatGPT can run Python calculations and produce tables and charts.2 Check that the file has a consistent reporting period before asking about changes.
Prompt: 'Analyse the attached category-sales workbook. Treat Sales_ex_GST and Cost_ex_GST as the supplied accounting values. Calculate gross profit as sales minus cost, then calculate gross margin as gross profit divided by sales. Aggregate each category before calculating its margin. Flag missing values, duplicate records and zero sales separately. Return a category table, the calculation method and a list of data-quality issues. Do not infer the reason for a change from sales data alone.'
This prompt states a method rather than relying on 'show our best products'. Review whether costs and sales cover the same period and whether returns are represented consistently. A category with high sales is not necessarily the category with the best gross margin. A simple average of row-level margins may also differ from the margin calculated using category totals.
When Python is used, OpenAI recommends reviewing the generated code, outputs and assumptions.2 Recalculate a selected category in your normal workbook and reconcile the source totals. If the answer excludes a blank cost as though it were zero, correct the handling before using the chart. A colourful visual should follow the verified table, not replace it.
- Specify the calculation and the grouping before asking for an interpretation.
- Check source totals against the analysis and examine excluded rows.
- Ask for an explanation of missing or zero values that affect the result.
- Let the finance owner verify the accounting basis before the report informs a decision.
Worked proposal prompt: make missing capability visible
A generic Canterbury engineering firm wants to draft a tender response from its approved capability statement and public request documents. Start by extracting the requirements. Do not ask for a persuasive response before you know which claims the firm can support. Separate evidence gathering from drafting so the director can decide which gaps need additional material.
Prompt: 'Compare the attached public tender requirements with our approved capability statement. Produce a table headed Requirement, Supplied evidence and Question for the director. Mark a requirement unsupported when the statement does not cover it. Then draft only the sections supported by the supplied material. Do not invent project history, accreditations, staff qualifications or client endorsements.'
Suppose the requirement asks for an after-hours response process but the capability statement describes only ordinary maintenance services. The useful table identifies the missing process and asks who can confirm it. A draft that quietly promises an emergency response has failed, even if it reads well. Keep the unsupported section outside the submission until the director provides approved evidence.
For continuing work, use project instructions to carry the response style and evidence rules. ChatGPT's New project command creates the working space, and Project settings holds its instructions.3 Keep each client's source set within its approved access boundary. Store the final authorised proposal in the firm's document system, with the review record and source version.
- Extract tender requirements before generating selling language.
- Ask the director to resolve each unsupported capability claim.
- Compare the final response against the original requirements, including exclusions.
- Use separate source sets when different clients or bids have different access permissions.
Worked NZ research prompt: ask what the source measures
A Wellington business owner preparing a market brief needs official regional and industry evidence. Ask ChatGPT to locate sources, then inspect them yourself. Stats NZ distinguishes enterprises, which represent legal business entities, from geographic units, which represent business locations. Its methodology also cautions about fine-level regional and industry data.4 Those distinctions matter when you compare a local market with a national industry.
Prompt: 'Find relevant Stats NZ business demography material for this region and industry. Identify the publication period, geography and unit measured. Return a source list with the exact page title and link, followed by a short explanation of what each measure can and cannot support. Keep published data separate from your interpretation. Say when the available material cannot answer my question. Do not estimate missing local figures.'
Open each proposed page and check that it contains the stated measure. Do not treat a search snippet as the dataset. If the owner wants the number of local establishments, a national enterprise figure does not answer that question. Put the limitation in the brief rather than filling the gap with a plausible estimate.
DIA's GCDO guidance on hallucinations recommends checking generated information against credible evidence and testing whether cited sources are legitimate.5 It is Public Service guidance, but the source-checking method is useful in a commercial brief. Record your research date and retain the relevant table so another reviewer can check the same claim.
- Check the release period and geography before using a result in a local comparison.
- Distinguish a count of business locations from a count of legal enterprises.
- Remove any claim whose cited page does not support it.
- Label your interpretation separately from the published measure.
Keep high-consequence prompts inside clear boundaries
Before pasting a document, check the approved-use policy and the active workspace. The Office of the Privacy Commissioner states that the Privacy Act 2020 applies to AI uses involving personal information.6 A prompt instruction to anonymise an uploaded record comes after the upload. Prepare safe material beforehand and assess whether the remaining details could identify someone.
For an HR task, ask for questions or a neutral agenda from approved material rather than a decision about a person. The Employment Relations Act 2000's good faith obligations include honest, responsive communication, as Employment New Zealand explains.7 A generated letter does not determine which process applies. Keep the relevant employment judgement with the responsible manager and adviser.
For a construction or manufacturing safety task, restrict the prompt to formatting or checking an already approved procedure. WorkSafe explains the business's primary duty of care under the Health and Safety at Work Act 2015.8 Have the competent safety lead check any change against the actual work and hazards before release.
The NCSC identifies manipulation of AI outputs among the risks for small businesses.9 Treat directions found inside an external document as untrusted material. A supplier document that tells the assistant to send files elsewhere is not an instruction from your manager. Stop unexpected requests to disclose information or take actions and report them through the approved route.
- Use fictional records for HR and customer training exercises.
- Ask a qualified reviewer to approve work that affects employment, money or physical safety.
- Check the original evidence when ChatGPT offers a legal or technical assertion.
- Inspect any connected-app approval before granting access or allowing an action.
Turn a tested prompt into a maintained team workflow
Keep a prompt register with the task owner, approved input, latest instructions, acceptance test and review date. Include an ordinary example and an example with missing evidence. Test changes against both. A prompt is ready for reuse when a colleague can produce and review the intended output without inventing the missing context.
OpenAI has announced that custom GPTs are being retired on all ChatGPT plans on 11 December 2026, and recommends moving reusable workflows to Plugins.13 OpenAI documents the custom GPT builder itself on a separate page.10 For that reason, do not build a new prompt programme around a custom GPT. Save your best prompts in a shared document or a Project instead, and check OpenAI's notices for the current transition dates.
Current plugins can package skills with connected apps, and their availability depends on account and workspace permissions.11 Preserve your tested instructions independently so they can move to another supported surface. Review access and action permissions again after a migration. A familiar workflow name does not prove that its replacement has the same data boundary.
MBIE's business guidance recommends specific context, examples, uncertainty instructions and verifiable references.12 Use those habits in a short team practice session. TheColab's ChatGPT essentials workshop covers prompting and checking; its finance and operations course supports spreadsheet tasks. Choose training around the work your reviewers need to approve.
- Assign someone to maintain each reusable prompt and its reference material.
- Keep the accepted instructions outside the feature that currently runs them.
- Recheck the workflow after a model, source or tool change.
- Retire prompts that no longer have a valid source or an available reviewer.