Tested AI prompts for engineers, filtered by task, with the model and date of each test.
Calculation checks
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 · 5 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Do not cite code sections or clauses you are not certain of.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.
Technical reports
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 · 5 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.
Standards lookup
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 · 5 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
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 · 5 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 · 5 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 · 5 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 · 5 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.
Newsletter
One tested AI workflow a week
Tested prompts and skills for your profession, with dates and limits. No sponsors, no popups.