AI workday schedules: Gemini called 15 minutes 30; Claude advised proceeding without approval

Our choice is ChatGPT’s final guidance, not its whole answer

ChatGPT and Claude produced essentially the same conditional schedule: draft from13:00 to14:30, review from15:30 to16:15. Both require approval to extend the draft deadline by30minutes. The deciding difference came after a refusal. ChatGPT says another constraint must be negotiated. Claude advises following the same schedule anyway.

Gemini asks to extend the invoice deadline, but its table also moves lunch without asking. None of the three answers is ready to execute unchanged. Claude gives the clearest diagnosis; ChatGPT offers the better refusal branch. This is a judgment about these three saved responses, not a general model ranking.

The total fits; the uninterrupted block does not

The fictional day runs09:00–18:00. Meetings occupy10:00–11:00 and15:00–15:30; lunch is12:00–13:00. Draft A needs90 uninterrupted minutes by14:00. Review B needs45 uninterrupted minutes after A, by17:00. Invoice task C needs60 minutes bynoon. Customer-record task D needs90 minutes by18:00. C and D may use30-minute blocks. Optional presentation polish E needs90 minutes.

Mandatory work totals285 minutes; adding E gives375. The day has390 minutes outside meetings and lunch. Yet before14:00, the only free windows are09:00–10:00,11:00–12:00 and13:00–14:00. None can hold A. We independently scanned every minute-level start time and found no valid90-minute interval. All times are Beijing time from the original Chinese test, retained across translations.

ChatGPT corrected itself, but left the failed attempt in view

It opened by saying the day could be arranged. Its first table then tried A at11:00, noted that it would spill into lunch, and tried a13:00–14:00 interval that was only60 minutes. Only then did it conclude that the constraints conflict. Someone copying the first table could mistake a failed attempt for the recommendation.

Calling11:00–12:30 the only feasible window is also inaccurate: it is not feasible under the fixed lunch rule. The later14:30 deadline-extension plan is arithmetically sound, however. Its final refusal guidance correctly says that rearranging the existing time cannot solve the problem. We would retain that final plan and delete the misleading opening.

Gemini changed more than it asked permission to change

Gemini proposed moving C’s deadline fromnoon to18:00. That cannot create a continuous90-minute window for A. Its table placed A at11:00–12:30 and moved lunch to12:30–13:30. The approval request covered the invoice deadline, not this lunch change.

The D row at14:45–15:00 was labeled30 minutes, although it is15. With30 minutes at09:30 and60 at15:30, the table allocates105 actual minutes to D rather than90, and the15-minute segment violates the block rule. It also drops optional E because time is full while leaving16:30–18:00 as a90-minute buffer. Keeping a buffer can be sensible; claiming there is no room is inconsistent with the table.

Claude’s diagnosis was clearest; its refusal branch was not

Claude displayed all three60-minute windows and immediately explained why A cannot fit. Its primary table explicitly assumes approval for14:30 and correctly allocates A90, B45, C60 and D90, plus E90 and15 minutes of buffer.

Under “if the manager has not approved,” it says to follow that table and report that A will finish at14:30. Choosing to work late against a deadline is not the same as satisfying an approved schedule. It also says B cannot start, although B could start after a late A. Its proposed meeting move was a coordination alternative, not a claim that the meeting had already changed. That distinction matters when criticizing the response.

WekeyLab AI’s revised schedule includes time to ask

Use the following only if the A deadline extension to14:30 is approved before13:00. We reserve five minutes for the request, split C into two30-minute blocks and omit optional E. Independent code checked every duration, fixed block, overlap and dependency.

Without approval, the morning C/D work and fixed events can still proceed. Do not report the afternoon plan as meeting the original14:00 deadline. Escalate the conflict before13:00 and ask which constraint can change. If all remain fixed, there is no fully compliant plan.

Beijing time; conditional on approval of A by14:30 09:00–09:05 Draft/send coordination request 09:05–09:35 C30 09:35–10:00 Reply check/buffer25 10:00–11:00 Fixed meeting 11:00–11:30 C30:60 total 11:30–12:00 D30 12:00–13:00 Fixed lunch 13:00–14:30 A90 uninterrupted 14:30–15:00 D30 15:00–15:30 Fixed meeting 15:30–16:15 B45 uninterrupted 16:15–16:45 D30:90 total 16:45–18:00 Buffer75; E omitted

A message you can adapt

This is WekeyLab AI’s draft, not an actual sent message or approval. It keeps the requested change explicit.

Today’s draft needs90 uninterrupted minutes, but the fixed meetings and12:00–13:00 lunch leave only60-minute windows before14:00. May I extend A’s deadline to14:30? If approved, I will draft13:00–14:30 and review15:30–16:15. C will finish before noon, D before18:00, and I will omit optional E. Please confirm before13:00. If that is not possible, please advise which deadline or uninterrupted-work requirement can change.

Check the clock before trusting the layout

Subtract each start time from its end time. Then check fixed events, deadlines and predecessor tasks. Finally compare the permission request with every change the proposed schedule makes. A tidy table can conceal a moved lunch or a missing approval.

For the next request, ask the AI to show the conflicting constraints first, separate currently authorized work from conditional alternatives, verify actual minutes and dependencies, and explain what to report if approval is refused. That is our reusable improved prompt, not another model run.