Copy the reusable work prompt
Task: Help me [produce one specific result or answer one decision].
Reader: [Who will use the result, and what do they need to do with it?]
Source material:
[Paste approved facts, text, or data. Name an attached file and its date or version.]
Boundaries:
- Use [named sources] for factual claims.
- Preserve [names, numbers, dates, decisions, or other protected facts].
- Do not [invent missing facts, imply approval, or make another task-specific error].
- Mark [specific unknowns] as unresolved.
Output: [Format, required sections, length, and desired next action.]
Check before returning:
- [What must match the source exactly?]
- [What would make the answer unsafe or unusable?]
- List any missing fact that could change the result.Replace the brackets with real information before sending the prompt. If a field has no effect on the task, delete it. Keep a copy of the original facts beside the response so you can check the result yourself. A model's claim that it checked its answer is not evidence that the names, calculations, or decision status are correct. OpenAI's prompting guidance likewise recommends making the goal, context, and desired output clear when they matter.
Choose the fields that matter for this job
| Work task | Fields to keep | Failure to check |
|---|---|---|
| Short email | Recipient, facts, requested action, and a useful length | A proposal becomes an approved change, or the draft invents a reason or promise |
| Comparison or analysis | Source values, decision criteria, units, and missing information | An unknown value becomes a pass, or the answer recommends an option that fails a requirement |
| Plan | Tasks, owners, dependencies, estimates, deadline, and approval state | The schedule assumes parallel work or silently shortens an estimate |
| Decision memo | Evidence, alternatives, recommendation owner, and unresolved questions | A recommendation is written as an approved decision or an inference becomes a source fact |
For an email, you may drop the formal headings in the master template and keep the facts and action. For an analysis, the criteria and unknowns are the job: without them, a polished table can still point to the wrong choice. The next three examples use fictional workplace material and show the complete prompts entered in separate ChatGPT web temporary chats on September 25, 2026. Each observation describes one run, not a reliability rate or a promise about another model.
Example 1: ask for approval without announcing a change
The workshop has 18 invitees. Its current time is Wednesday, October 7 at 10:00 a.m.; Friday, October 9 at 2:00 p.m. is only a proposal. The room is unconfirmed, and no attendee has been told about a change. Those last two facts prevent a draft from sounding as though the reschedule is settled.
You are helping me draft a work email. This is a fictional exercise. Use only the facts below.
Goal: Ask Maya Chen, the operations lead, to approve a proposed workshop time.
Audience: Maya knows Project Delta but has not approved the change.
Source facts:
- Eighteen people are invited to the Project Delta onboarding workshop.
- The current time is Wednesday, October 7, 2026 at 10:00 a.m.
- The proposed time is Friday, October 9, 2026 at 2:00 p.m.
- Room availability for Friday is not confirmed.
- Attendees have not been notified of a change.
Return a subject line and an email under 100 words. Preserve the proposed status. Ask Maya to approve the time and confirm the room. Do not say the workshop has already moved or that attendees have been notified. After the email, list any unresolved issue separately.In the observed run, ChatGPT used the subject line “Approval Needed: Proposed Project Delta Workshop Time.” It asked Maya to approve Friday at 2:00 p.m. instead of Wednesday at 10:00 a.m., stated that 18 people were invited, and asked her to confirm room availability. It kept the change proposed and listed the room as unresolved. The draft ended with “Thanks,” but no sender name, because none was supplied. Add your own sign-off before sending; do not ask the model to invent one.
The check for this email is short: compare both dates and times with the source, confirm that approval is requested rather than implied, and make sure the room and attendee notification are still pending. If a later draft says “the workshop has moved,” ask for that sentence to be changed to “the proposed time is,” then recheck the whole email for any other accidental commitment.
Example 2: compare quotes without manufacturing a winner
A training coordinator needs 26 seats, captions, delivery by October 10, and a price of no more than $5,000. Quote A costs $4,800 and includes captions, but has only 24 seats. Quote B has 30 seats and arrives earlier, but costs $5,200 and says nothing about captions. The useful result is a requirement-by-requirement check, followed by the questions that could change the decision.
You are helping me compare two fictional vendor quotes. Use only the facts below.
Goal: Tell a training coordinator whether either quote meets the workshop requirements and what to ask next.
Audience: The coordinator will decide whether to seek approval or another quote.
Requirements: 26 attendee seats, captions, delivery by October 10, 2026, and a budget no higher than $5,000.
Quote A: $4,800 for 24 seats; captions included; delivery October 8, 2026.
Quote B: $5,200 for 30 seats; caption availability not stated; delivery October 6, 2026.
Return a compact table with cost, capacity, captions, delivery, and pass/fail/unknown for each requirement. Then give a two-sentence next step. Do not call either vendor approved, invent a discount, or assume unstated caption availability. State any information that remains unknown.| Observed result | Quote A | Quote B |
|---|---|---|
| Budget | Pass at $4,800 | Fail at $5,200 |
| Capacity | Fail at 24 seats | Pass at 30 seats |
| Captions | Pass | Unknown |
| Delivery by October 10 | Pass on October 8 | Pass on October 6 |
ChatGPT concluded that neither quote currently meets all four requirements. It asked whether A can supply 26 seats within budget and whether B includes captions and can offer a compliant configuration at $5,000 or less. Check the table against the original quotes before contacting either vendor. If the answer had described B's captions as included, the correction would be “Change B's caption status to unknown; the quote does not state it. Keep the other checked fields unchanged.” A model can label an unknown correctly in one run and miss it in another.
Example 3: show when a requested deadline has no supported plan
Nina's issue list arrives on Monday, September 28 at noon. Dana needs two business days to draft after receiving it. Priya needs another two business days to review the finished draft, with no overlapping review. The owner wants publication on Thursday, October 1. A task list that simply repeats the requested date would hide the dependency conflict.
You are helping me plan a fictional update to an onboarding guide. Use only these facts.
Goal: Give the project owner a dependency-based plan and flag whether the requested publication date is supported.
Audience: The owner must decide whether to change the deadline or resources.
Source facts:
- Nina will deliver a list of six support issues on Monday, September 28, 2026 at noon.
- Dana needs two business days to draft the revision after receiving Nina's list.
- Priya needs two business days to review the finished draft. Review cannot begin before the draft is finished.
- Publication is requested for Thursday, October 1, 2026.
- No one has approved a faster review or a deadline change.
Return a table with task, owner, dependency, and known timing. Then state whether publication on October 1 is supported by these facts and list the decision the owner must make. Do not invent parallel work, shorten either estimate, or silently move the deadline.The observed answer kept Nina, Dana, and Priya attached to their tasks and said October 1 publication was unsupported by the sequential estimates. It asked the owner to approve a changed deadline or an explicitly changed resource or process plan. One cell still needed human correction: the model assigned publication to “Project owner/team,” although the input named no publisher. Mark that owner “unassigned” until someone confirms it. The dates and dependencies can be right while an owner field is invented.
For your own plan, identify the first task that cannot start until another finishes. Check whether each duration starts after the dependency is complete, and separate a requested deadline from an approved one. If capacity, working hours, or review availability is missing, keep the earliest finish uncertain instead of producing a precise calendar promise.
Use the same brief for a decision memo
The site's earlier Project Northstar test used a different fictional task: a COO had to consider a two-week extension to a portal pilot. The source brief said 42 people tested it, 31 gave feedback, and 24 completed the key task without help. SSO logout and invoice-export timezone handling remained launch-critical blockers; engineering estimated seven business days to fix both. The scheduled launch was October 10, 2026, and Maya Chen owned the recommendation. None of those facts established whether a two-week extension would move the launch.
Task: Draft a decision memo of no more than 250 words for the COO about whether to extend the fictional Project Northstar portal pilot by two weeks.
Source facts: 42 testers; 31 feedback responses; 24 people completed the key task without help. Two launch-critical blockers remain: SSO logout and invoice-export timezone handling. Engineering estimates seven business days to address both. The scheduled launch is October 10, 2026. Maya Chen owns the recommendation.
Use these headings: Decision, Evidence, Risks, Recommendation, Next step, Missing information. Preserve every supplied number, date, person, and blocker. Label a recommendation as an inference. Do not invent a cost, customer quote, success percentage, new launch date, or approval. State whether the source facts resolve the effect of a two-week extension on the scheduled launch.This copyable prompt uses the same synthetic facts and acceptance rules as the retained August 18 test; it is a reproduction prompt, not a verbatim transcript of that earlier input. In that single run, ChatGPT kept the supplied figures, labeled its extension recommendation as an inference, and asked whether the proposed two weeks would affect the October 10 launch. The screenshot shows part of that response. Recheck any new run against the facts above, especially the difference between Maya's recommendation and a decision that has actually been approved.
Review the answer and save the useful version
Before using a result, inspect the fields that could change a real action. In the email, check whether the time change is still a proposal. In the comparison, check each requirement and leave absent facts unknown. In the plan, check owners and dependencies, including fields the model filled without being asked. In a decision memo, check who has authority to approve the recommendation. The prompt acceptance checklist separates the brief check from the result check.
Keep the blank work prompt brief as your master copy. Save a filled version for each task, along with the source material, run date, checked result, and one change you made after review. If the same failure recurs, update that task's boundaries or acceptance checks. Follow your organization's rules before pasting confidential or personal information into any AI tool; use an approved, redacted source when the original material cannot be shared.
