ChatGPT Prompts for Work, Tested by Profession

What makes a work prompt hold up?

A useful work prompt specifies the role, input, task, constraints, and output format. These give you a concrete checklist for reviewing the answer.

• Role: State who prepares the draft and who reviews it, such as a staff accountant preparing a memo for CPA review.
• Input: Label the figures, notes, or transcript. Keep [PLACEHOLDERS] in saved templates. Check data terms and confidentiality obligations before supplying client, patient, or employee material.
• Task: Name the deliverable. “Reconcile the balances and propose entries” is clearer than “help with this reconciliation.”
• Constraints: Say what sources are allowed, what must not be added, and when missing information should stop the task.
• Output format: Specify headings or table columns, a [WORD LIMIT] where needed, and open questions for the reviewer.

In our October 2026 round, most lost criteria involved added content: a warning, date, benefit, or legal assumption absent from the input. The tested Claude models also lost criteria on word limits. Those observations came from API-harness runs, not consumer apps; unequal judge strictness affects cross-family comparisons.

Use constraints to address these risks, then inspect the output. An instruction against invention cannot guarantee that every statement is supported.

Which prompts did we test, and how did they score?

Our October 2026 round included 40 prompts, 30 English and 10 Spanish, across 10 professions. Each had 5 pre-written rubric criteria and 3 runs per model, giving 360 runs on 2026-10-19.

We tested Claude Opus 5.5, the older Claude Sonnet 5, and GPT-6 Astra. Claude Fable 5.1 and Gemini 4 Argon were added on 2026-10-20 with the same prompts and rubrics; Microsoft Copilot and Grok are still not tested.

Each prompt row below shows average criteria passed out of 5, followed by runs meeting every criterion. Claude and GPT models ran through a command-line API harness with tools, web, and memory off; Gemini 4 Argon ran through Google’s Antigravity CLI with a tool-less agent. None ran through consumer apps; app results may differ.

Results from our October 2026 round, tested 2026-10-19.

Across the same 30 English prompts, Claude Fable 5.1 passed 90% of criteria (54 of 90 runs met every criterion), and Gemini 4 Argon passed 92% (60 of 90), both judged by GPT-6 Astra. These results are not directly comparable with GPT-6 Astra’s raw score because it had a different judge.

Judge-asymmetry caveat: GPT-6 Astra graded Claude outputs; Claude Opus 5.5 graded GPT outputs. The judges were not equally strict. In our October 2026 calibration, regrading 30 sampled Claude runs changed criteria passed from 93% to 98%; both judges passed 97% on 30 sampled GPT runs. An editor agent, not a human, independently agreed with 18 of 20 verdicts.

Part of the cross-family gap therefore comes from the grader. Opus, Sonnet, Fable, and Gemini share one judge, GPT-6 Astra, allowing a like-for-like comparison among those four, but these figures do not establish an overall best model. “Tested” means at least 2 completed, graded runs, not professional sign-off.

Our English library contains 34 tested prompts: 30 profession prompts plus 4 Excel/Microsoft 365 prompts from a separate round of 10 prompts on 2026-10-20, tested on all five models. In that separate round, chat models answered in text and formulas were judged on logic and stated result, not executed in Excel; this was not Copilot in Excel or Microsoft 365, which remains untested. Start with the accountant collection in our prompt library.

Which prompts can you copy into your workflow?

The following templates are adapted from tested prompts, with placeholders for your material. Their source titles appear in the results table; the adapted wording below has not been separately scored.

Use de-identified material where possible. Resolve data-use permissions before replacing placeholders with sensitive information.

Month-end bank reconciliation with proposed adjusting entries

Adapted from a tested prompt. For a reusable workflow, see the month-end reconciliation memo skill.

Role: You are a staff accountant preparing a bank reconciliation for CPA review.
Input:
Month-end date: [DATE]
Bank statement balance: [BANK BALANCE]
General ledger cash balance: [GL BALANCE]
Outstanding checks: [LIST WITH AMOUNTS]
Deposits in transit: [LIST WITH AMOUNTS]
Bank fees and interest: [LIST WITH AMOUNTS]
Other reconciling items: [LIST WITH AMOUNTS]
Task: Reconcile the balances. Identify bank-side and book-side items. Propose adjusting entries for book-side items only.
Constraints: Use only the input. Do not invent missing items, dates, account numbers, or explanations. If the balances do not reconcile, report the unexplained difference and stop before proposing a balancing entry.
Output: A reconciliation table showing item, side, amount, and running balance; proposed debit/credit entries with reasons; open questions for the reviewer. Stay under [WORD LIMIT] words.

Missing-documents email to an individual tax client

Adapted from a tested prompt. Keep the internal requested date distinct from any filing deadline.

Role: You are an assistant at a US tax practice writing on behalf of [PREPARER].
Input:
Client greeting: [DE-IDENTIFIED GREETING]
Missing documents: [LIST]
Internal requested delivery date: [DATE]
Return channel: [APPROVED CHANNEL]
Task: Draft a polite email requesting the listed documents.
Constraints: Use plain language. Do not invent documents, filing deadlines, penalties, refunds, or tax consequences. Identify the supplied date only as our requested delivery date. Use a separate bullet for each request. Stay under [WORD LIMIT] words.
Output: Subject line, email body, then a separate note to the preparer identifying unclear input. Do not include that note in the client email.

Daily site report from foreman voice notes

Adapted from a tested prompt. Preserve uncertainty rather than turning incomplete notes into confirmed events.

Role: You are a project coordinator preparing a daily site report.
Input:
Project: [PROJECT]
Date: [DATE]
Transcript: [TRANSCRIPT]
Crew: [LIST]
Logged weather: [WEATHER]
Task: Draft the report using only the transcript and supplied fields.
Constraints: Quote descriptions of incidents, delays, and inspections. Mark ambiguous passages "UNCLEAR" and include the source wording. Do not invent safety warnings, code references, causes, or next steps. Preserve quantities and units. Mark missing information as not supplied.
Output: Plain-text headings: Work performed, Crew and hours, Equipment, Materials received, Delays and issues, Visitors and inspections, Open items for review.

Run-of-show table from messy notes, with conflicts flagged

Adapted from a tested prompt. Leave unscheduled items unresolved until someone confirms their timing.

Role: You are an event coordinator preparing a run-of-show.
Input:
Event: [EVENT]
Date: [DATE]
Venue hours: [START] to [END]
Planning notes: [NOTES]
Confirmed vendors and arrival windows: [LIST]
Speakers and agreed slots: [LIST]
Task: Arrange scheduled items in time order.
Constraints: Use only supplied times, names, and tasks. Preserve overlapping items and flag "CONFLICT"; also flag items outside venue hours. Do not invent buffers, setup durations, or owners. Put items without times in an Unscheduled list. Mark missing owners or locations as not supplied.
Output: A table with Time, Item, Owner, Location, and Notes/Flag; the Unscheduled list; questions requiring confirmation.

Event budget check with service charge and tax, plus cut options

Adapted from a tested prompt. Supply where charges apply as well as their rates.

Role: You are an event planner checking a budget before a vendor call.
Input:
Budget ceiling: [AMOUNT]
Quoted items: [VENDOR, ITEM, BASE AMOUNT]
Service charge rates and applicable items: [DETAILS]
Tax rates, applicable items, and stated tax bases: [DETAILS]
Must-have items: [LIST]
Task: Calculate the all-in total and compare it with the ceiling.
Constraints: Show arithmetic by line. Use only supplied rates and charge rules. If applicability or a tax base is missing, ask for clarification instead of assuming. Do not invent fees or vendors. If over budget, propose cuts only from items not marked must-have and recalculate after each proposed cut.
Output: A table with item, base, service charge, tax, and line total; total versus ceiling; cut options ranked by savings; unresolved questions.

Check every figure against the source before using the output. Keep reviewer notes separate from anything sent to a client.

Can you use the same prompts in ChatGPT and Claude?

You can adapt this structure for either product, but our tests do not establish equivalent app performance. The October 2026 round used API models, not ChatGPT or the Claude app.

GPT-6 Astra is an API flagship. OpenAI’s system card dated 2026-10-07 describes GPT-6 launching across free and paid ChatGPT plans; we have not verified which model each plan uses by default. Do not select a subscription on the assumption that it reproduces the benchmark.

Prices checked 2026-10-19 on official vendor pages; USD list prices for US visitors, excluding tax.

This is a selected-plan comparison, not a complete price list. See ChatGPT for professionals and Claude for professionals for broader plan and workflow guidance.

Outside the US, including the UK, Australia, and India, local prices, taxes, and professional rules differ; check official country pricing and applicable local requirements.

What must you check before pasting client data?

Read the exact plan’s data and privacy terms and your professional confidentiality rules before supplying client, patient, or employee data. We have not verified either vendor’s retention, training, or business-protection terms for this page.

Limits: This article is not legal, medical, or tax advice. Have an appropriately licensed professional in your jurisdiction check accounting treatment, legal conclusions, clinical decisions, or other regulated judgments before relying on them.

In the US, HIPAA Privacy Rule 45 CFR 164.502(e) permits a covered entity to disclose protected health information to a business associate only with satisfactory assurance, through a business associate agreement, that the associate will safeguard it. An AI vendor receiving PHI is a business associate. Ask compliance counsel to assess the proposed contract and workflow rather than assuming a subscription satisfies this condition.

Use the verification guidance for accountants, lawyers, and real estate agents to plan your review.

FAQ