The verify rule — three categories of trust

::what the AI returned

Good list. Real uses, not hypotheticals. Here's the split:

**TRUST (used it raw, fine to keep doing that):**
- #2 cold email draft — voice/style work, you're the final reader, low blast radius
- #6 LinkedIn bio rewrite — same category, you read it before posting
- #8 Spanish translation — short, conversational text; you can sanity-check the vibe even without fluent Spanish. (If it were a legal notice from the contractor, this moves to VERIFY.)
- #10 Midjourney hero image — aesthetic output, you decide if it looks right with your own eyes

**VERIFY (you should have checked before acting, may not have):**
- #1 PDF summary for client meeting — this is the dangerous one. 40 pages compressed to a summary in 20 min means you walked into a client meeting trusting a model's pick of what mattered. At minimum: skim the exec summary + any section the AI flagged + anything with numbers. Going forward, paste the summary back and ask "what did you leave out that a sharp client might ask about?"
- #4 Postgres MRR query — SQL runs. Run it on a known slice and check the number against something you already know (last month's total, a single customer). MRR queries famously double-count upgrades or miss pauses.
- #7 California cap gains rate — Perplexity is closer to OK here because it cites, but you still click the cite. Tax rates move and models confidently quote old ones.
- #9 apartment lease clause — Claude is decent at this but "explain" can drift into "advise," and lease language is jurisdiction-specific. Cross-check with the actual statute or a 10-min call to a tenant hotline if the clause matters to money.

**DON'T USE (wrong tool for the job):**
- #3 Stripe founding year told to coworker as fact — you stated a year you didn't verify. Stripe was founded in 2010, not 2009. This is the textbook bad pattern: low-stakes-feeling fact, zero friction to check, you became the source of a wrong claim. One Wikipedia tab would have caught it.
- #5 "best pre-workout for someone over 35" — health + supplements + personalized recommendation = three failure modes stacked. Models will confidently name brands and dosages without knowing your meds, BP, caffeine tolerance, or what "best" even means to you. Ask a pharmacist or your doctor; use AI to prep questions, not to pick the product.

**The pattern in your 10:**
You trust AI correctly when you're the final filter (writing, images, casual translation). You over-trust it when speed pressure is on (#1, #3) and when the topic feels small but the consequence is real (#3, #5). The verify rule isn't "check everything" — it's "check when a wrong answer costs more than the 60 seconds it takes to check."

One concrete change for next week: before quoting any AI-produced fact to another human, open one tab and confirm. That single habit kills most of the bad uses on this list.