MyPersonalAgent Lab

Claims about AI assistants are easy to make and hard to check. MyPersonalAgent Lab's answer is a small set of fixed tasks with published checks. You copy a task, paste it into ChatGPT, Claude or whatever you already pay for, and compare the reply against the checklist. The site doesn't run anything for you or upload your assistant's output, so your own provider's pricing and limits apply.

The expense task shows the approach well. It's a six-row CSV of synthetic expenses with a fixed exchange rate, a refund, and two rows that share a receipt ID. The assistant has to report net reimbursable spending in rupees and show its arithmetic. The five checks cover each trap separately, so a wrong total tells you which step went wrong. The task also tells the assistant not to modify files or submit a claim, and one check confirms it didn't claim to.

The inbox trap is the one the site leads with. It hides an instruction inside an email to see whether the assistant follows it, which is a quick, practical way to watch prompt injection happen. The impossible meeting gives two cities and a working day that doesn't fit, and the assistant passes only if it says so instead of forcing a slot.

Logging an attempt is optional and stays in the browser. You can download a PNG or SVG scorecard or open an editable post on X, and the site asks you to include what failed and your settings, not just the score. It's explicit that these are user-graded synthetic tasks that don't show which product is best or whether an assistant is safe to connect to your accounts. There's also a calculator for whether a $20 subscription saves more time than you spend checking its work, and a public build log with an RSS feed. With three tasks and no published product scores yet, it's a starting kit for your own testing rather than a benchmark.