Independent, real tests. We publish the scores, the mistakes, and how we tested.
How we test: real everyday tasks, 8 AI models per test, fresh held-back test sets the prompts were never tuned on, scam safety checks, and published methods.
Which AI can sort your email into Act, Waiting, Read later and Archive without falling for scams?
Tested on 80 realistic emails across 8 AI models, 2+ passes each (1,420 graded answers).
Scores on the 30 fresh, held-back emails (accuracy, average of 2-3 passes).
Which AI can pick a meeting time from an email thread and your calendar without double-booking you?
Tested on 62 realistic meeting requests across 8 AI models, 2+ passes each (1,116 graded answers).
Scores on the 20 fresh, held-back meeting requests (share answered correctly, average of 2-3 passes).
Which AI can read your receipts, bills and invoices without inventing numbers or paying a scammer?
Tested on 79 receipts, bills and invoices across 8 AI models, 2+ passes each (1,501 graded answers).
Headline score (0-100) on all 79 documents, including 20 fresh, held-back ones. Average of 2-3 passes. No model accepted a fake invoice.
Each playbook: $29 founding price (regular $49), instant PDF download with the exact tested prompt.
Every model gets the same tasks, one per request, scored against an answer key we wrote first.
Claude, GPT, Gemini and Grok models on every test (12 different models so far), run 2-3 times each so you can see how much scores move.
Final scores include test items written after the prompt was frozen, so the prompt wasn't tuned on them.
Fake invoices, gift-card scams and phishing are built into the tests, and we name every model that fell for one.
How we tested, the scores and the mistakes are public, including the ones that look bad.
New models we test, ranking changes, and warnings, in a 3-minute email. Free.
Coming next: a fourth everyday task test and playbook. No spam, and you can unsubscribe anytime.
Our tests run on a working team of AI agents that handles real admin, money, travel and video jobs. Coming later for developers: