Recommendation letters and reviews, written honestly
Lesson 5 of 6 · Running a group
1 · Learn the move · Verified-data override
A recommendation letter is a signed factual document, and the override rule keeps it one: the model drafts only from your evidence file - projects with dates, skills you observed, results with their actual scope - and every superlative it cannot ground gets flagged, not smoothed in. 'Top student I have trained' is a sentence you either mean or should not sign; the draft's job is to make that choice visible, not to make it for you. Peer review runs under a harder rule: the manuscript in your queue is confidential, and pasting it into a tool may violate journal and funder policy - NIH prohibits AI in its peer review, and venues differ. Check the venue's current policy; when unclear, the manuscript stays out and the model sees only your own notes. Flag and refer, never assume.
Draft a recommendation letter for [person] for [position] from this evidence file ONLY: [projects with dates, observed skills, results with scope]. Rules: every claim traces to a file line; superlatives I did not evidence get flagged [UNSUPPORTED: your call], not smoothed in; comparison claims ('among the best') flagged the same way. For review work: I paste only MY OWN notes, never the manuscript - and remind me to check the venue's current AI policy before even that.2 · Your turn. You write the prompt
Two letters due Friday - your best student in five years, and a solid-but-not-exceptional one applying to the same program - plus a review you owe a journal by Monday. The temptation on letter two is inflation; the temptation on the review is pasting the PDF into the tool at 11 p.m. You want all three done honestly.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.