Learn AI from an AI tutor, ready when you are.

Detail is exactly what a language model is good at producing. It has seen thousands of real citations, so it knows the pattern: plausible author names, a journal that fits the field, a sensible year. When it lacks a real source, it can generate one that fits the pattern perfectly. Formatting tells you nothing about existence. So treat every citation as unverified until you find it in a library database or the journal itself. Try it now: search for the exact title. What do you find?

Neither reply tells you much. Models lean towards agreeing with you, so both the praise and the agreement are partly echo. Get feedback with something to check against instead. Open a new chat, paste the email, and ask: 'List the three weakest points for a busy hiring manager, quoting the exact sentence for each. No praise.' Then judge each point yourself. Before you do that, write down one thing you want the hiring manager to remember after reading. That becomes your first criterion.

It will sometimes, so design for that. Add a fourth category, Unsure, for anything that does not clearly fit, and send it to a review folder you check daily. Label every email the automation touches so a misfiled message is still easy to find. Before switching it on, run it over fifty past emails and compare with what you would have chosen. If complaints get mixed up with admin, sharpen the one line definitions. How would you define a complaint in one sentence?

That is join fan out. Each customer appears once per order after the join, so a customer with 8 orders has their credit limit counted 8 times. Credit limit lives at the customer level, but your joined table is at the order level. Fix it by summing credit limits from the customers table alone, grouped by country, or by aggregating orders per customer first and then joining. Quick check: count rows in customers, then in your joined result. What numbers do you get?