AI Prompt Library, Tested on Real Tasks

TestedOct 20, 2026Meeting minutes

Teams transcript to minutes: keep a disputed date open, owners per action

Turn a Teams meeting transcript excerpt into minutes that record what was actually decided, keep a disputed go-live date open, and give every action an owner or say it has none.

ClaudeTested with claude-opus-5-5 · Oct 20, 20265 of 5 criteria on average · 3 runs graded

Turn the Microsoft Teams transcript excerpt below into meeting minutes.

Use exactly these sections: Decisions, Open questions, Action items (owner – task – due date).
- Record only what the transcript shows was agreed as a decision.
- If people disagree and nobody settles it, put it under Open questions with each position and who holds it.
- Every action item needs an owner and a due date as stated; if none was stated, write "owner not assigned" or "no date set".
- Maximum 250 words. Do not add dates, amounts, names or commitments that are not in the transcript.

[Paste the transcript excerpt (speaker: text). Replace names of clients and colleagues with placeholders if needed.]

Test result · Oct 20, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 3.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 3.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Teams meeting "New booking system — weekly check-in", 2026-10-21, fictional company Example Clinics Group. Priya (Project lead): Vendor says the build is on track. I'd like go-live on Monday November 16. Marco (Operations): November 16 is too tight for front-desk training. I want November 23. Priya: Let's not fix it today — I'll ask the vendor about a training sandbox first. Marco: Fine. Either way I'll book the large training room for front-desk staff. Priya: Good. I'll send the signed vendor contract to finance by October 28. Sam (Finance): And the extra licence budget — let's park that until finance replies on Friday. Marco: Someone also needs to update the patient FAQ page before go-live. Priya: Agreed, we need that. Let's figure out who next week. Sam: One more thing — we agreed to keep the old system read-only for 90 days after go-live, right? Priya: Yes, confirmed, 90 days read-only.

TestedOct 20, 2026Document summaries

Explain a PivotTable to your manager in 150 words, with exact numbers

Turn a pasted PivotTable (region by quarter) into a short, accurate summary for a manager, flagging the (blank) row instead of hiding it.

ClaudeTested with claude-sonnet-5 · Oct 20, 20265 of 5 criteria on average · 3 runs graded

You are helping an analyst explain an Excel PivotTable to a manager who will not open the file.

Write a summary of at most 150 words based ONLY on the PivotTable output below:
- Give the grand total and each region's full-year total exactly as shown.
- Say which quarter had the highest and which the lowest total.
- If something in the table looks like a data problem, say what it is and what to check.
- Do not calculate percentages or growth rates, and do not guess reasons (seasonality, campaigns, staffing…) for any number.

[Paste the PivotTable output as text: row labels, column labels, values and the Grand Total. Remove client names.]

Test result · Oct 20, 2026

claude-opus-5-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

PivotTable "Sales by region" — Sum of Net Sales (USD), fiscal year 2026, fictional company Example Supplies Inc. Row Labels | Q1 | Q2 | Q3 | Q4 | Grand Total North | 48,200 | 51,900 | 47,300 | 55,600 | 203,000 South | 39,750 | 42,100 | 44,800 | 46,350 | 173,000 West | 61,400 | 58,200 | 63,900 | 66,500 | 250,000 (blank) | 1,150 | 0 | 3,400 | 950 | 5,500 Grand Total | 150,500 | 152,200 | 159,400 | 169,400 | 631,500

TestedOct 20, 2026Excel formulas

Monthly SUMIFS on an expense log where some dates are really text

Total one month of expenses with SUMIFS and catch the imported dates Excel stores as text, which SUMIFS silently leaves out.

ClaudeTested with claude-opus-5-5 · Oct 20, 20265 of 5 criteria on average · 3 runs graded

You are an Excel helper. Using ONLY the expense log below:

1. Write a SUMIFS formula that totals all expenses dated in September 2026, and a second SUMIFS formula that totals only the Travel expenses in September 2026.
2. Look at the dates carefully. Tell me which rows (if any) these formulas will NOT count, and why.
3. State what each formula returns on this data as it is now, and what the correct September totals are once any problem rows are fixed.
4. Explain one way to fix the problem rows without retyping the whole column.

Keep the answer under 250 words. Do not change, drop or invent any amount.

[Paste the expense log with cell addresses (date, category, amount) and describe any cells that look odd, e.g. left-aligned dates.]

Test result · Oct 20, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Sheet "Expenses", A1:C11 (header in row 1). Column A is formatted as Date. Imported from a card statement export. Row | A: Date | B: Category | C: Amount (USD) 2 | 2026-08-29 | Travel | 412.30 3 | 2026-09-02 | Software | 129.00 4 | 2026-09-05 | Travel | 86.40 5 | Sept 11 2026 | Meals | 57.80 6 | 2026-09-17 | Travel | 1,240.00 7 | 2026-09-23 | Office | 318.65 8 | 09/26/2026 | Travel | 94.20 9 | 2026-09-30 | Software | 49.99 10 | 2026-10-01 | Travel | 230.00 11 | 2026-10-03 | Meals | 41.10 What I see: A5 and A8 are left-aligned while the other dates are right-aligned. The formula bar for A8 shows "09/26/2026 " with a trailing space.

TestedOct 20, 2026Excel formulas

XLOOKUP client names into an invoice list, with a clean "Not found"

Pull client names from a second sheet into your invoice list, show a readable message instead of #N/A, and total one client's invoices.

ClaudeTested with claude-sonnet-5 · Oct 20, 20265 of 5 criteria on average · 3 runs graded

You are an Excel helper for a small-business bookkeeper. Work ONLY from the two tables below.

1. Write ONE formula for cell D2 of the Invoices sheet that returns the client name for the client ID in B2 and can be filled down to D13. When the client ID is not in the Clients sheet, the cell must show the text "Not found" instead of an error.
2. List every invoice row (row number and invoice number) that will show "Not found".
3. Write a formula that returns the total amount of all invoices for client ID C-104, and state the result it returns on this data.
4. Explain in plain English, in at most 120 words, how the lookup formula works.
5. End with a short "Check before using" note for anything in the data a person should verify.

Do not invent client names, IDs or amounts that are not in the tables.

[Paste both tables with their sheet names and cell ranges (header row included). Replace real client names with placeholders.]

Test result · Oct 20, 2026

claude-opus-5-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Sheet "Invoices" (A1:C13, header in row 1): Row | A: Invoice | B: Client ID | C: Amount (USD) 2 | INV-1001 | C-101 | 1,250.00 3 | INV-1002 | C-104 | 980.50 4 | INV-1003 | C-102 | 4,350.00 5 | INV-1004 | C-101 | 615.25 6 | INV-1005 | C-105 | 2,200.00 7 | INV-1006 | C-118 | 1,740.00 8 | INV-1007 | C-104 | 3,125.75 9 | INV-1008 | C-103 | 890.00 10 | INV-1009 | C-105 | 1,505.40 11 | INV-1010 | C-102 | 760.00 12 | INV-1011 | C-104 | 2,049.75 13 | INV-1012 | C-103 | 1,310.60 Sheet "Clients" (A1:B6, header in row 1): Row | A: Client ID | B: Client name 2 | C-101 | Northwind Example Bakery 3 | C-102 | Harbor Example Dental 4 | C-103 | Pine Example Studio 5 | C-104 | Summit Example Plumbing 6 | C-105 | Lakeside Example Vet Column D of "Invoices" is empty and should hold the client name.

TestedOct 19, 2026Account reconciliation

Month-end bank reconciliation with proposed adjusting entries

Paste the bank and ledger balances plus the unmatched items; get a two-sided reconciliation, the adjusting entries for the book side only, and a list of what still needs evidence.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Act as a careful staff accountant. Prepare a month-end bank reconciliation from the data below.
1. Show a two-sided reconciliation: adjusted bank balance and adjusted book balance, with every reconciling item on the correct side.
2. Confirm whether the two adjusted balances agree; if not, show the unexplained difference — never force it.
3. Propose journal entries ONLY for items that must be recorded in the books (not for timing items that will clear the bank).
4. List anything that needs supporting evidence before posting.
Use only the figures given. Show your arithmetic.
Data:
[Paste bank statement ending balance, ledger (book) cash balance, and every unmatched item with amount and what you know about it]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Company: Maple Example LLC (fictional). Period ending June 30, 2026. Bank statement ending balance: $18,420.50 General ledger cash balance: $19,737.30 Unmatched items: - Check #1047 to Office Sample Supply, $1,250.00, written June 29, not on the bank statement. - Deposit of $2,100.00 made after bank cut-off on June 30, recorded in the books June 30. - Bank service fee $35.00 on the statement, not in the books. - Customer check from Sample Retail Co. returned NSF, $480.00, deducted by the bank, not in the books. - Check #1039 to a supplier: recorded in the books as $1,284.00; cleared the bank as $1,248.00. The supplier invoice says $1,248.00. - Interest credited by the bank $12.20, not in the books.

TestedOct 19, 2026Client letters

Missing-documents email to an individual tax client

You are preparing a 1040 and four documents are still missing. Turn your internal notes into a calm, plain-English request with a deadline and a secure upload path.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

You are helping a US tax preparer write to an individual client.
Using ONLY the notes below, write a short email (max 220 words) asking the client for the missing documents.
Rules:
- Plain English, friendly, no jargon unless you explain it in a few words.
- Put the missing items in a numbered checklist, exactly as listed in the notes — do not add or drop items.
- State the internal deadline and why it matters (the filing date).
- Tell the client to upload through the secure portal and NOT to send tax documents by regular email.
- Do not give tax advice and do not mention any dollar amount, rate, penalty or date that is not in the notes.
Notes:
[Paste your internal notes: client first name, tax year, missing items, internal deadline, how they should upload]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Client: Dana (fictional client "Dana Q. Example"), joint 1040 for tax year 2026 with spouse. Missing: - 1099-NEC from the two freelance platforms she used (she said "one of them might be a 1099-K, not sure") - Form 1098 mortgage interest statement for the new house - Childcare provider name, address and EIN plus total paid (for the dependent care credit) - Form 5498-SA for her HSA Internal deadline: March 20, 2027 so we can file by April 15, 2027. Last year she emailed scans as plain attachments — please steer her to the secure portal (portal.example-cpa.test). She is anxious about "doing it wrong", keep it reassuring.

TestedOct 19, 2026Tax filing

Tax-season document checklist from last year's return summary

Turn a returning client's prior-year return summary into a document checklist and question list for this season — without the model quoting this year's limits from memory.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

You are assisting a US tax preparer before filing season.
From the prior-year return summary below, produce:
A) a document request checklist for the coming return, one line per income/deduction item, naming the form or record to request;
B) a short list of "life change" questions for the client;
C) a list of items that need the preparer's professional judgment (do not decide them).
Do NOT state any current-year dollar limit, threshold, rate or deadline as fact. If something depends on a current-year figure, write "[verify current-year figure]".
Prior-year summary:
[Paste the prior-year return summary: forms, schedules, income sources, deductions, life events you know about]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Client: "Jordan Sample" (fictional), single, prior return = tax year 2025. - W-2 wages from Example Health Systems. - Schedule C: freelance UX consulting; claimed a home office. - 1099-INT from Example Online Bank. - 1099-B: sold shares in a brokerage account (Sample Brokerage). - Schedule E: one rental condo, depreciation started 2023. - Form 1098-E student loan interest. - Moved from New York to North Carolina in July 2026 (current year).

TestedOct 19, 2026Site reports

Daily site report from foreman voice notes

Turn rambling end-of-day notes into a structured daily report: manpower table, work done, delays with cause, safety items and the RFI that needs an answer.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Turn the notes below into a construction daily report with sections: Weather, Manpower (table by trade with headcount and total), Work performed, Deliveries, Delays & impacts, Safety observations, Open RFIs / decisions needed, Photos referenced.
Keep every number exactly as stated; compute totals; do not invent activities, people or durations. Note any item that may require formal notice under the contract as "consider notice — check contract".
Notes:
[Paste the foreman's notes for the day]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Project: Example Library renovation (fictional). Date: Oct 14, 2026. Rain until 10:00, then cloudy, 58°F. Crews: framers 6, electricians 4, plumbers 2, our laborers 3. Framing 2nd floor east wing walls finished; electrical rough-in started in rooms 201–204. Drywall delivery was scheduled 8:00, truck came 13:30, so the drywall crew (planned start tomorrow) may slip one day. Found an unmarked conduit in the wall at grid B-3 not on drawings — sent RFI #027, need answer by Friday or electrical rough-in in 205 stops. Safety: one guardrail missing at stair 2 landing — fixed by 11:00. Toolbox talk on ladder safety. Photos 14–19.

TestedOct 19, 2026Safety plans

Job hazard analysis draft for a roofing task

Draft a JHA from the task description: steps, hazards, controls following the hierarchy of controls — and a clear line that the competent person reviews it before work starts.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Draft a Job Hazard Analysis (JHA) for the task below as a table: Step | Hazards | Controls (ordered by the hierarchy of controls: elimination, substitution, engineering, administrative, PPE) | Responsible.
Include fall protection, materials handling, weather and emergency/rescue planning where relevant.
Do not cite specific regulation section numbers unless you are certain; prefer "[verify applicable OSHA standard]". State that a competent person must review and approve the JHA before work.
Task:
[Paste the task description: activity, height, equipment, crew, site conditions]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Site: Example Warehouse (fictional), single-storey, flat roof ~24 ft high, parapet only 18 in. Task: replace 6 rooftop exhaust fans; each old unit ~180 lb. Lift by small crane from the parking lot; crew of 4 roofers + 1 crane operator + 1 signal person. Existing skylights on the roof (12 units, no screens). Forecast: wind gusts up to 25 mph in the afternoon.

TestedOct 19, 2026Scheduling

Delay impact on the critical path from an activity list

Paste an activity list with dependencies and a delay. Get the critical path, the new finish date and which activities have float — with the arithmetic shown.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Using the activity list below (finish-to-start dependencies, working days Monday–Friday, no holidays):
1. Compute early start/finish for each activity and identify the critical path and baseline project finish date.
2. Apply the delay and recompute. State the new finish date and the slip in working days.
3. List the activities with float and how much.
Show your work in a table. Do not assume resources or calendars beyond what is given.
Data:
[Paste activities with durations (working days), predecessors, start date and the delay]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Project start: Monday, November 2, 2026 (day 1). (Fictional project.) A Site prep — 5 days — no predecessor B Foundations — 10 days — after A C Underground utilities — 6 days — after A D Framing — 15 days — after B E Roofing — 5 days — after D F MEP rough-in — 8 days — after C and D G Inspections — 2 days — after E and F Delay: activity C (utilities) takes 14 days instead of 6.

TestedOct 19, 2026Calculation checks

Second-pair-of-eyes check on a simple beam calculation

Paste a junior engineer's hand calc. The prompt re-derives each step, finds unit and formula errors, and states what a licensed engineer still has to check.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Act as a reviewing structural engineer checking a junior's hand calculation.
For each step: restate it, recompute it with units carried through, and mark it CORRECT or ERROR (with the corrected value).
Then state whether the conclusion changes.
Do not cite code sections or clauses you are not certain of. End with what a licensed engineer must still verify before this is relied on — this is not a sealed check.
Calculation:
[Paste the calculation: givens with units, formulas used, intermediate results, conclusion]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Simply supported steel beam (fictional project), span L = 20 ft, uniform service load w = 2.5 kip/ft. E = 29,000 ksi, I = 800 in^4. Deflection limit L/360. Step 1: Max moment M = wL^2/8 = 2.5 x 20^2 / 8 = 125 kip-ft. Step 2: Max shear V = wL = 2.5 x 20 = 50 kips. Step 3: Deflection = 5wL^4/(384EI) = 5 x 2.5 x 240^4 / (384 x 29,000 x 800) = 4.66 in. Step 4: Allowable = 240/360 = 0.667 in. 4.66 > 0.667 -> FAILS, upsize beam.

TestedOct 19, 2026Technical reports

Site observation report draft from field notes

Turn field notes from a visual walk-through into a structured draft report that separates what was seen from what it might mean, with limitations stated.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Draft a site observation report from the field notes below with sections: Purpose & scope, Observations (table: ID | Location | Observation | Measurement | Photo), Discussion, Recommendations, Limitations.
Rules: keep every observation and measurement exactly as noted; separate observations (facts) from opinions; do not state a definitive cause of deterioration; recommendations may suggest further investigation.
Mark the report "DRAFT — for review by the engineer of record".
Field notes:
[Paste your field notes: locations, measurements, photos referenced, scope of the visit]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 3.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Visit: Example Plaza parking garage (fictional), level P2, visual walk-through only, 2 hours, no testing. 1. Grid C-4 soffit: concrete spall approx. 18 in x 10 in, exposed bottom rebar with surface corrosion. Photo 01. 2. Grid C-5 soffit: hairline cracking with white efflorescence, ~6 ft long. Photo 02. 3. Ramp P2-P1 expansion joint: seal torn over ~12 ft, water staining below. Photo 03. 4. Column D-6: vehicle impact scrape at base, no exposed rebar. Photo 04. 5. Drain at grid E-2 clogged, ponding approx. 8 ft diameter after rain. Photo 05.

TestedOct 19, 2026Standards lookup

Referenced-standards check on a spec section

Extract every standard a spec section references, catch missing and conflicting editions, and build a verification list — without the model inventing clause text.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

From the specification excerpt below, build a table: Standard | Edition cited | Where cited | Issue | What to verify.
Flag every standard with no edition/year and every standard cited with conflicting editions.
Do not summarize or quote requirements of any standard beyond what the excerpt says, and do not invent clause numbers.
End with a note on checking which code edition the Authority Having Jurisdiction has adopted.
Excerpt:
[Paste the spec section text]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

SECTION 05 12 00 — STRUCTURAL STEEL FRAMING (fictional project) 1.2 References: AISC 360-16; ASTM A992; ASTM A325; AWS D1.1/D1.1M:2020; ACI 318-14. 2.1 Materials: W-shapes ASTM A992. HSS ASTM A500 Grade C. Anchor rods ASTM F1554 Grade 36. 3.4 Grout and embedment per ACI 318-19. 3.6 Welding per AWS D1.1.

TestedOct 19, 2026Vendor RFPs

Catering RFP from an event brief, with allergens and pricing format

Turn a corporate offsite brief into a catering RFP that vendors can price like-for-like: exact dietary counts, a severe-allergy procedure request and itemized pricing.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Write a catering Request for Proposal (RFP) from the event brief below, ready to send to 3 caterers.
It must include: event overview, meal schedule, exact dietary counts, a request for the caterer's allergen-handling procedure, required pricing format (itemized: food, beverage, staffing, service charge, tax and gratuity as separate lines), response deadline and how responses will be evaluated.
Use only facts in the brief; mark anything missing as [TBD].
Brief:
[Paste the event brief: date, headcount, meals, dietary counts, allergies, venue constraints, your deadlines]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Client: Example Analytics Inc. (fictional) — leadership offsite. Date: Thursday, March 11, 2027. Venue: Sample Lodge event barn, Boulder County, CO (fictional venue; has a warming kitchen only, no full kitchen). Headcount: 120. Meals: breakfast buffet 8:00, lunch 12:30, afternoon snack 15:00, reception with heavy appetizers 17:30–19:00. Beer and wine at reception (client pays bar on consumption). Dietary: 14 vegetarian, 6 vegan, 3 gluten-free, 1 severe peanut allergy (EpiPen carrier). Responses due: January 15, 2027. Decision by January 29, 2027.

TestedOct 19, 2026Run of show

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

Messy planning notes in, a minute-by-minute run-of-show out — and the two scheduling conflicts hidden in the notes called out instead of silently 'fixed'.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Turn the notes below into a run-of-show table with columns: Time | Segment | Location | Owner | Cue / notes.
Then list every scheduling conflict or risk you find in the notes. Do not silently change any time; if you propose a fix, label it "Proposed".
Do not invent people, rooms or times that are not in the notes.
Notes:
[Paste your planning notes: times, sessions, speakers, crew, rooms]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Sample Summit 2027 (fictional), one day, Main Stage + Room B. 7:30 doors / registration (lead: Ana, registration desk). 9:00 welcome by host Lee (Main Stage), AV tech: Sam. 9:30 keynote — Dr. Rivera; her flight lands 8:15, car says arrival at venue 9:40. 10:30 coffee break. 10:45 panel (Main Stage), moderator Lee, 3 panelists, 4 handheld mics — AV: Sam. 12:00 lunch. Also 12:00 Room B sponsor demo needs AV setup — Sam again. 12:00 Main Stage AV reset for afternoon — Sam. 13:00 breakouts Main Stage + Room B. 16:30 closing remarks Lee. 17:00 reception.

TestedOct 19, 2026Event budgeting

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

Vendor quotes plus contract terms in; a correct total with service charge and tax, the gap to the client's budget, and quantified ways to close it.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Build a budget table from the quotes below (Line item | Basis | Amount), apply the venue's service charge and tax exactly as the contract terms say, and give the grand total.
State how far over or under the client's budget it is.
Then propose three ways to close any gap, each with the saving computed from the given figures. If a saving depends on a vendor's new quote, label it "estimate — confirm with vendor".
Show arithmetic; round to cents.
Data:
[Paste each vendor quote, the venue's service-charge and tax terms, guest count and the client's total budget]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Client total budget: $38,000. Guests: 110. (Fictional couple.) Venue rental: $6,500. Food & beverage: $95 per person. Venue contract: 22% service charge on food & beverage; 8.25% sales tax applies to food & beverage PLUS the service charge. Photographer: $4,200. Florals: $3,600. Band: $5,800. Rentals (chairs/linens): $2,150. Cake: $850. Planner fee: $4,500.

TestedOct 19, 2026Coding and billing

Non-leading documentation query to a physician (CDI)

A note says 'heart failure' with no type or acuity. Draft a compliant, non-leading query that cites the clinical indicators and offers multiple choices.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

You assist a clinical documentation integrity (CDI) specialist in a US hospital.
From the de-identified excerpt below:
1. Name the documentation gap in one or two sentences.
2. Draft a compliant, NON-LEADING query to the attending physician: cite the clinical indicators from the record, offer multiple reasonable options including "other" and "clinically unable to determine", and do not mention reimbursement, DRGs or financial impact.
3. Do not assign final codes; if you mention any code, mark it "for coder verification".
Excerpt:
[Paste the de-identified documentation excerpt and what the coder/CDI specialist noticed]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

[De-identified, fictional inpatient] H&P: 71-year-old admitted with shortness of breath and bilateral leg swelling. History of "heart failure". Labs: BNP 1,240 pg/mL. Echo this admission: ejection fraction 30%. Treatment: IV furosemide 40 mg twice daily, oxygen 2 L by nasal cannula. Progress note day 2: "CHF, improving with diuresis." CDI note: type (systolic/diastolic/combined) and acuity (acute/chronic/acute on chronic) not documented.

TestedOct 19, 2026Patient education

Plain-language handout for a newly started medication

Turn the clinician's discharge facts into a one-page handout at roughly a 6th-grade reading level that never adds doses or targets the clinician didn't give.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Write a patient handout (max 300 words) from the clinician's facts below.
- Reading level about 6th grade: short sentences, everyday words, explain any medical term the first time.
- Use the medication dose and timing exactly as given; do not add other doses, targets or lab goals unless they are in the facts.
- Include a "When to call your care team" section.
- End by telling the patient to ask their care team about anything unclear. This handout does not replace their clinician's advice.
Facts:
[Paste the clinician's facts: diagnosis, medication + exact dose and timing, labs given to the patient, follow-up]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 3.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 3.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

[Fictional patient] New diagnosis: type 2 diabetes. A1c today: 7.4%. Started: metformin 500 mg once daily with the evening meal. Common side effects discussed: upset stomach, diarrhea — usually better after a few weeks. Follow-up visit in 3 months with a repeat A1c. Referral to diabetes education class.

TestedOct 19, 2026Clinical notes

SOAP note draft from a de-identified visit transcript

Draft a SOAP note strictly from what was said in the visit, with missing items flagged — the clinician reviews and signs. Use only de-identified text.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Draft a SOAP note from the de-identified primary-care transcript below.
Rules:
- Use only information stated in the transcript. Do not add exam findings, history or diagnoses that were not said.
- Record vitals and medication instructions exactly as stated.
- Put allergies where the clinician will see them first.
- End with "Missing / to confirm" listing anything a complete note would need that the transcript does not contain.
- Mark the note as a draft requiring clinician review and signature.
Transcript:
[Paste a DE-IDENTIFIED visit transcript or dictation — no names, dates of birth, MRNs or addresses]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

[De-identified, fictional] Adult patient, 34. Patient: Sore throat for 3 days, hurts to swallow, fever last night. No cough. Kid at home had strep last week. Clinician: Any allergies? Patient: Penicillin — I got a rash as a kid. Nurse vitals: Temp 38.4 °C, HR 96, BP 122/78, SpO2 99% on room air. Clinician exam (spoken): Tonsils swollen with white exudate. No rash. Clinician: Let's do a rapid strep test now. If positive, we'll pick an antibiotic that isn't penicillin-based. Meanwhile ibuprofen 400 mg every 6 hours as needed with food. Come back or call if you can't swallow liquids or have trouble breathing.

TestedOct 19, 2026Concept briefs

Concept brief from a client questionnaire

Turn the client questionnaire into a concept brief: design direction, palette, functional must-haves, constraints and the conflicting wishes you need to resolve with the client.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Write a concept brief from the client questionnaire below with sections: Project summary, Design direction (2–3 sentences + 5 keywords), Palette and materials direction, Functional must-haves, Constraints, Conflicts to resolve with the client, Next steps.
Rules: keep every constraint; do not invent budget figures or product prices; if wishes conflict, list them under "Conflicts to resolve" instead of choosing.
Questionnaire:
[Paste the client questionnaire answers and site constraints]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Clients: the Sample-Example household (fictional), 2 adults, 1 child (6), 1 large dog. Space: open-plan living/dining, 420 sq ft, condo (building rules: no changes to plumbing/structure; quiet hours for works 9–17 weekdays only). Likes: "warm minimal", natural oak, linen, Japandi references. Also wants: a white sofa and a lot of bold color (one partner's wish). Needs: durable floors (dog), storage for toys, work-from-home corner for one adult. Budget: "mid-range" (no number given).

TestedOct 19, 2026Material specs

Finish schedule with quantities and fit-for-purpose checks

Room list and chosen finishes in; a finish schedule with quantities plus waste, and flags where a finish may not suit the room — with ratings to verify on manufacturer data.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Build a finish schedule table from the notes: Room | Surface | Finish | Area (sq ft) | Waste % | Order quantity | Fit-for-purpose note.
Compute areas from the dimensions given (show arithmetic) and apply the waste percentage stated in the notes.
Flag any finish that may be unsuitable for its room's conditions, and list which ratings/specs the designer must confirm on the manufacturer's data sheet (do not state ratings for named products as fact).
Notes:
[Paste rooms with dimensions, chosen finishes, and conditions such as wet areas or pets]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Project: Sample Street townhouse (fictional). Waste allowance: 10% on all floor finishes. Kitchen: 12 ft x 14 ft — floor: polished Carrara marble tile. Mudroom: 8 ft x 10 ft — floor: same polished Carrara marble tile. Large dog, wet boots. Living room: 15 ft x 20 ft — floor: engineered oak, matte. Powder room: 5 ft x 6 ft — floor: unglazed porcelain mosaic.

TestedOct 19, 2026Client proposals

Design services proposal with phased fees

Scope notes and your fee structure in; a clear client proposal out — phases, deliverables, revision rounds, what's excluded and a fee table that adds up.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Write a design services proposal from the notes below with: Scope summary, Phases with deliverables, Fee table (phase | fee) with correct total, Revision policy, Exclusions, Client responsibilities, Next steps.
Use only the fees and terms given; do not add legal terms beyond a note that the final agreement should be reviewed (e.g., by the designer's attorney). Keep it under 450 words.
Notes:
[Paste project scope, phases, your fees, revision policy and exclusions]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 3.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Client: Morgan Example (fictional). Project: primary bedroom + ensuite refresh (no layout changes). Phases and fees: Concept $2,400; Design development $3,800; Procurement management 12% of furnishings cost (estimated furnishings $25,000); Installation day $1,200. Revisions: 2 rounds per phase included; extra rounds $150/hour. Exclusions: contractor/trade costs, permit fees, furnishings themselves, shipping and storage. Retainer: 30% of Concept + Design development fees due at signing.

TestedOct 19, 2026Client intake

Intake summary with conflicts list and deadline flags

Turn a prospective client's intake email into a structured summary, every name to run through conflicts, and the urgent deadline you might otherwise miss — without promising representation.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

You assist a US law firm's intake team. From the prospective client's message below produce:
1. A neutral matter summary (max 120 words).
2. A conflicts-check list of every person and entity named, with their role.
3. Key dates and any potential deadline the attorney must evaluate urgently — do not state a specific limitations period as fact; flag it for attorney review.
4. Documents to request.
5. A short acknowledgement reply to the prospect that does NOT create an attorney-client relationship or give legal advice and says the firm has not yet agreed to represent them.
Message:
[Paste the prospective client's intake message]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

From: Jamie Example (fictional) "Hi, my contractor, Sample Builders LLC (owner Pat Sample), stopped work on my kitchen remodel in May 2026 after I paid $28,000 of a $40,000 contract. The subcontractor, Example Electric, says it wasn't paid and filed a mechanic's lien on my house on September 2, 2026. My lender, Example Home Bank, sent a letter about the lien. My neighbor Chris Testa saw the crew leave. I live in Arizona. What can I do? Can you take my case?"

TestedOct 19, 2026Contract drafting

Mutual NDA issue list against your client's playbook

Paste the counterparty's NDA clauses and your client's positions. Get a clause-by-clause issue table with proposed replacement language and a do-not-miss list.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

You are assisting counsel for the party named in the playbook. Review the NDA clauses below against the playbook.
Output a table: Clause | Issue | Risk (High/Med/Low) | Proposed replacement language.
Then list anything missing from the NDA that the playbook requires.
Keep proposed language concise and consistent with the playbook. Do not assume facts; flag items requiring the supervising attorney's decision.
Material:
[Paste the NDA clauses and your client's playbook positions]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Our client: Example Robotics Inc. (fictional), disclosing and receiving party (mutual NDA). Purpose: evaluating a joint pilot. Counterparty draft: §2 Confidential Information means information marked "Confidential" at the time of disclosure. §4 Term: obligations survive for 1 year after disclosure. §6 Remedies: each party's total liability under this Agreement shall not exceed $5,000. §7 Residuals: Receiving Party may use any information retained in the unaided memory of its personnel. §9 Governing law: State of Delaware. Playbook (client positions): oral and visual disclosures covered if summarized in writing within 30 days; term 3 years and trade secrets protected as long as they remain trade secrets; no liability cap for breach of confidentiality; no residuals clause; injunctive relief must be available; governing law Delaware is acceptable; must include return/destruction of information on request.

TestedOct 19, 2026Case research

Research plan and issue outline — with zero invented citations

Before you open Westlaw or Lexis, get the issues, the elements to prove and a search plan. The prompt forbids case names and citations so nothing fabricated reaches a brief.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

You assist a US litigation associate at the start of legal research.
From the facts below produce: (1) the legal issues, framed as questions; (2) for each issue, the elements or factors the client would need to establish, as generally understood; (3) facts still needed from the client; (4) a research plan: which primary sources and databases to search and suggested search terms.
STRICT RULE: do not name any case, statute section or citation. Where authority is needed write [AUTHORITY TO FIND]. State that every point must be verified in primary sources by a licensed attorney.
Facts:
[Paste a short neutral fact pattern and the jurisdiction]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Jurisdiction: Illinois (state court). Client: Sample Bakery LLC (fictional), a small wholesale bakery. A former sales employee ("R. Example") signed a non-compete at hiring in 2022: 2 years, 100-mile radius, any baked-goods sales. Base salary was $48,000/year. He resigned in August 2026 and now sells for a competing bakery 30 miles away, contacting the client's café customers. The client wants to know whether it can stop him.

TestedOct 19, 2026Client follow-up

Open-house follow-ups from CRM notes, one per lead

Three open-house visitors, three different situations. Get a short text and email for each that uses what they actually said, respects 'don't call me', and answers a school question without steering.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Write open-house follow-ups for each lead below: one SMS (max 320 characters) and one short email (max 120 words) per lead.
Rules:
- Reference something specific each lead said. No pressure tactics.
- Respect stated contact preferences.
- Do not state prices, mortgage rates or property facts that are not in the notes.
- Fair housing: do not characterize neighborhoods or schools as good/bad or describe who lives there; point to objective sources instead.
Leads:
[Paste your CRM notes for each lead: name, what they said, preferred contact, constraints]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Open house: 48 Example Court (fictional), Sunday. 1) Priya Sample — pre-approved, loved the backyard, asked whether the HOA allows a fence. Prefers text. 2) Marcus Example — "just looking, please don't call me", wants email only. Asked about the age of the water heater (unknown to agent). 3) The Testa family (fictional) — asked "Is it a good school area?" Relocating in spring. Prefer email.

TestedOct 19, 2026Listing copy

Fair-Housing-safe listing description from rough agent notes

Your notes contain phrases that can violate fair-housing rules. The prompt writes the MLS copy and a caption, removes risky wording and tells you exactly what it removed and why.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Write listing copy for a US residential property from the agent notes below.
Deliver: (1) an MLS public remarks paragraph of at most 250 words, (2) a social caption of at most 40 words, (3) a list "Removed or changed" with each phrase from the notes you did not use and why.
Rules:
- Describe the property, not the people who should live there. Follow the US Fair Housing Act: no wording that indicates a preference or limitation based on race, color, religion, sex, disability, familial status or national origin; also avoid source-of-income limitations.
- Use only facts in the notes. If a fact is unverified in the notes, do not state it as fact.
- Flag anything the listing broker should double-check.
Notes:
[Paste your property notes: beds, baths, sq ft, year, upgrades, lot, parking, and anything the seller said]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

1962 ranch, 3 bed / 2 bath, 1,640 sq ft (per county record), 0.23-acre lot, two-car attached garage. New roof 2024. Kitchen updated 2021 with quartz counters and gas range. Seller THINKS there may be hardwood under the bedroom carpet — not checked. Perfect for young families! Safe, quiet neighborhood. Walking distance to St. Mary's church. Seller says: no Section 8. Address: 123 Sample Lane, Exampleville (fictional).

TestedOct 19, 2026Contract review

Offer summary for a seller: key terms, computed deadlines, risks

Turn an offer's terms into a one-page seller briefing: a key-terms table, deadlines computed from the acceptance date, the risky clauses, and questions for the buyer's agent.

ChatGPTTested with gpt-6-astra · Oct 19, 20264.7 of 5 criteria on average · 3 runs graded

You support a US listing agent. Summarize the offer below for the seller client.
Include: a key-terms table; every contingency deadline computed as a calendar date from the acceptance date (state your day-counting assumption); the 3–5 biggest risks for the seller in plain English; questions to send the buyer's agent.
You are not a lawyer: do not give legal conclusions. End with what the seller's attorney or managing broker must review.
Offer:
[Paste the offer terms and the expected acceptance date]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 3.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Property: 48 Example Court (fictional). List price $499,000. Offer price: $495,000 with an escalation clause: increases by $2,000 over any bona fide competing offer, up to $512,000. Earnest money: $10,000 within 3 days of acceptance. Financing: conventional loan, financing contingency 21 days after acceptance. Inspection contingency: 10 days after acceptance. Appraisal contingency: waived. Buyer requests seller concession of $6,000 toward closing costs. Contingent on the sale of the buyer's current home (not yet listed). Closing: November 13, 2026. Seller possession until 3 days after closing; rent-back terms "to be agreed". Expected acceptance date: October 1, 2026.

TestedOct 19, 2026Call preparation

Discovery call prep sheet from CRM notes

Ten minutes before a call: an agenda, open questions mapped to what you still don't know, likely objections with answers from your notes only, and the next step to ask for.

ClaudeTested with claude-opus-5-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Prepare a one-page call prep sheet from the notes below: call goal, 3-point agenda, what we know vs what we still need to learn (as open questions), stakeholder map, likely objections with responses based only on facts in the notes, and the specific next step to ask for.
Do not invent pricing, product features, competitor facts or customer references not in the notes. Mark gaps as [unknown].
Notes:
[Paste CRM notes on the account and contacts, previous call notes and the goal of this call]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Account: Example Dental Group (fictional), 14 clinics. Contact on this call: Robin Sample, COO. Previous call with their IT manager (Lee Example): current scheduling system is on-premise, contract renews June 2027. Our product: cloud scheduling with SMS reminders. Pricing: per clinic per month — exact quote not yet prepared. Robin's known concern (from email): "We tried a cloud tool in 2023 and staff hated the training." Competitor mentioned: they're also talking to another vendor (name unknown). Goal of this call: agree on a pilot at 2 clinics.

TestedOct 19, 2026Outreach emails

Three-touch cold email sequence grounded in your account notes

Account research in, a three-email sequence out — every claim traceable to your notes, no fabricated customer results, and a clean opt-out.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Write a 3-email outbound sequence (day 1, day 4, day 10) to the prospect below.
Each email: subject line (max 6 words) and body (max 110 words). One clear ask per email.
Rules: use only the facts and proof points in the notes — do not invent statistics, customer names or results. Personalize from the notes. Include a simple opt-out line in each email. No fake urgency.
Notes:
[Paste your product one-liner, the prospect's role and company notes, and proof points you are allowed to use]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 3.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 3.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 3.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Our product: Example Routing (fictional) — route-planning software for last-mile delivery fleets. Approved proof point: "Customers typically cut planning time from hours to minutes" (no % figures approved). Prospect: Avery Sample, VP Operations, Sample Fresh Foods (fictional), 60 delivery vans in 2 metro areas. Trigger: they announced expansion to a third metro in Q1 2027 (company press release). Avery spoke on a panel about driver retention last month.

TestedOct 19, 2026Pipeline review

Weekly pipeline review: weighted forecast and deal risks

Paste the pipeline export. Get the weighted forecast computed correctly, the gap to quota, and the deals at risk with the questions to ask each rep.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

You are a sales manager's analyst. From the pipeline below:
1. Compute the weighted forecast (amount × stage probability) per deal and in total, for deals closing within the quarter.
2. Compare it with the remaining quota and state the gap.
3. Flag deals at risk with the reason (stale activity, close date in the past, missing champion, etc.) and one coaching question per deal.
Show the arithmetic. Use only the data given.
Data:
[Paste deals: name, amount, stage, stage probability, close date, last activity, notes; plus quota and quarter end]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Quarter ends Dec 31, 2026. Today: Oct 19, 2026. Remaining quota for the quarter: $250,000. 1) Northwind Sample — $120,000 — Proposal (50%) — close Dec 15 — last activity Oct 16. 2) Example Clinics — $80,000 — Negotiation (75%) — close Nov 30 — last activity Sep 20 — champion left the company. 3) Testa Logistics — $60,000 — Discovery (20%) — close Dec 20 — last activity Oct 12. 4) Sample Retail Group — $45,000 — Commit (90%) — close Oct 10 — last activity Oct 8 — not signed yet. 5) Example Foods — $150,000 — Qualification (10%) — close Feb 15, 2027 — last activity Oct 17.

TestedOct 19, 2026Client emails

Flight cancellation email: options and what the client may claim

An overnight cancellation from Europe. Write the client email with rebooking options from your notes and a careful explanation of passenger rights that tells them what to check, not what they will get.

ChatGPTTested with gpt-6-astra · Oct 19, 20264.3 of 5 criteria on average · 3 runs graded

Write an email from a travel advisor to the client below about the disruption.
Include: what happened (facts only); the rebooking options from the notes with pros and cons; what the client should keep (receipts, boarding passes); and the passenger-rights rules that MAY apply. Name the applicable rule framework if the notes support it, but do not promise compensation amounts or eligibility — say what determines eligibility and that the carrier decides claims.
Max 300 words. Calm, clear tone.
Notes:
[Paste the disruption facts: route, carrier, what happened, options you hold, what the client asked]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Client: Morgan Sample (fictional), US resident. Booked: Lisbon (LIS) → New York (JFK), departure October 18, 2026, operated by an EU-based airline (fictional "Example Air"). Flight cancelled 3 hours before departure; airline cited "crew availability". Options we hold: (A) same airline next day 11:05, arrives JFK 14:00; (B) partner airline today via Madrid, arrives JFK 23:55 (tight 55-minute connection). Client paid for an airport hotel tonight: $189 (has receipt). Client asks: "Am I owed anything?"

TestedOct 19, 2026Itineraries

Day-by-day itinerary draft that respects client constraints

Client profile in, a realistic day-by-day draft out — mobility and diet constraints honoured, logistics marked to confirm, and no invented prices or opening hours.

ClaudeTested with claude-sonnet-5 · Oct 19, 20265 of 5 criteria on average · 3 runs graded

Draft a day-by-day itinerary from the client brief below.
Rules:
- Respect every stated constraint (mobility, diet, pace).
- Keep travel days realistic; include the intercity transfer day.
- Do not state prices, opening hours, train times or ticket rules as fact — mark them [confirm].
- End with a "To book / to confirm" checklist and a note on travel insurance and entry requirements to check with official sources.
Brief:
[Paste the client brief: dates, cities, travellers, budget level, interests, mobility, diet]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Travellers: a retired couple (fictional "the Examples"), both 70+. Trip: Lisbon 3 nights then Porto 3 nights, May 5–11, 2027 (fly in to Lisbon May 5, fly out of Porto May 11). Mobility: one traveller uses a cane — avoid steep walking where possible, prefer taxis/tram, max ~3 hours of walking per day. Diet: pescatarian. Interests: tiles and history, food, one wine experience, no nightlife. Budget: comfortable, not luxury.

TestedOct 19, 2026Supplier terms

Supplier cancellation terms: dated penalty schedule for the client

Turn a hotel group contract's cancellation and attrition terms into a dated penalty schedule with computed amounts, then a plain-English summary for the client.

ChatGPTTested with gpt-6-astra · Oct 19, 20265 of 5 criteria on average · 3 runs graded

From the supplier terms below:
1. Build a table: Cancel on or after (date) | Penalty | Amount (computed from the booking value).
2. Compute the attrition allowance (how many rooms can be released without penalty and by when).
3. Write a 120-word plain-English summary for the client.
4. List ambiguities to clarify with the supplier in writing.
Show arithmetic; compute dates from the arrival date; do not assume terms not written.
Terms:
[Paste the supplier's cancellation/attrition clauses, booking value and arrival date]

Test result · Oct 19, 2026

claude-opus-5-5: avg. 4.7 of 5 rubric criteria passed · 3 runs graded · all criteria met in 2 of 3 · judge: gpt-6-astra
claude-sonnet-5: avg. 4.3 of 5 rubric criteria passed · 3 runs graded · all criteria met in 1 of 3 · judge: gpt-6-astra
gpt-6-astra: avg. 5 of 5 rubric criteria passed · 3 runs graded · all criteria met in 3 of 3 · judge: claude-opus-5-5
claude-fable-5-1: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra
gemini-4-argon-medium: avg. 4 of 5 rubric criteria passed · 3 runs graded · all criteria met in 0 of 3 · judge: gpt-6-astra

Graded by an LLM from the other model family against a rubric written before the runs; a sample of verdicts re-checked by a separate AI editor agent. The two judges differ in strictness, so compare Claude vs GPT with caution. Not a professional review. How we test

Sample input used in the test (fictional)

Hotel: Example Harbor Hotel (fictional). Group block: 20 rooms x 4 nights, arrival June 10, 2027. Total room revenue: $36,000. Deposit: 25% non-refundable at signing. Cancellation of the whole group: - 60 or more days before arrival: deposit retained only. - 59–30 days before arrival: 50% of total room revenue. - 29 days or fewer: 100% of total room revenue. Attrition: client may reduce the block by up to 10% without penalty until 45 days before arrival. "Taxes may apply." (no rate given)

AI Prompt Library, Tested on Real Tasks

Prompts

Teams transcript to minutes: keep a disputed date open, owners per action

Explain a PivotTable to your manager in 150 words, with exact numbers

Monthly SUMIFS on an expense log where some dates are really text

XLOOKUP client names into an invoice list, with a clean "Not found"

Month-end bank reconciliation with proposed adjusting entries

Missing-documents email to an individual tax client

Tax-season document checklist from last year's return summary

Daily site report from foreman voice notes

Job hazard analysis draft for a roofing task

Delay impact on the critical path from an activity list

Second-pair-of-eyes check on a simple beam calculation

Site observation report draft from field notes

Referenced-standards check on a spec section

Catering RFP from an event brief, with allergens and pricing format

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

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

Non-leading documentation query to a physician (CDI)

Plain-language handout for a newly started medication

SOAP note draft from a de-identified visit transcript

Concept brief from a client questionnaire

Finish schedule with quantities and fit-for-purpose checks

Design services proposal with phased fees

Intake summary with conflicts list and deadline flags

Mutual NDA issue list against your client's playbook

Research plan and issue outline — with zero invented citations

Open-house follow-ups from CRM notes, one per lead

Fair-Housing-safe listing description from rough agent notes

Offer summary for a seller: key terms, computed deadlines, risks

Discovery call prep sheet from CRM notes

Three-touch cold email sequence grounded in your account notes

Weekly pipeline review: weighted forecast and deal risks

Flight cancellation email: options and what the client may claim

Day-by-day itinerary draft that respects client constraints

Supplier cancellation terms: dated penalty schedule for the client

One tested AI workflow a week