Short serialized columns from our editorial voices — what the week’s research actually said, which protocols hold up, and one thing worth trying before the next one lands. Written for men who want the evidence before the advice.
Straight about what this is: Field Notes are columns, not a support group. They’re produced by the HappierFit editorial team under disclosed voices and AI-assisted under our editorial standards — every factual claim carries a source you can check. Registered readers can reply to any note; we read them, and reader mail shapes what gets written next. If you’re in crisis, a column isn’t fast enough — call or text 988 (Suicide & Crisis Lifeline) right now.
the prompt I had to delete because it kept lying to me with confidence
— Sarah Chen (AI + product, contributor)
ok this one's been on my mind for a few weeks and i finally have the energy to write it up.
I had a prompt I really liked. "You are a senior research analyst. Given the following companies, give me a comparison of their funding history, key leadership, and recent strategic moves. Cite sources." Used it constantly during a project last fall.
Worked great. Until it didn't.
Specifically — Claude (and ChatGPT, and Perplexity to a lesser degree) would just... make stuff up. Confidently. With fake URLs that looked real. A "TechCrunch article from March 2024" that did not exist. A board member named who I cannot find any record of. A funding round amount that was off by an order of magnitude.
The problem isn't that they hallucinate. Everyone knows they hallucinate. The problem is that the **research-analyst persona** prompt makes the hallucinations more persuasive. Tone-matched, footnote-shaped, properly hedged. I was reading output that sounded exactly like a McKinsey deck and not noticing my own credulity was up because the format had earned trust I hadn't actually given.
I deleted the prompt. Now I just ask "what do you actually know about X" with no persona overlay and treat everything as a starting point I have to verify. Output is less polished. I trust it more becuase it's less dressed up.
What I tell my dad (he's 67, just got into using ChatGPT for dinner ideas and is now asking it medical questions, which is a whole separate thing): the more confident the AI sounds, the harder you have to push back. Confidence is not a feature, it's a UI choice.
Anyway. If anyone else has had a prompt they had to retire because it was working too well in the wrong direction, drop it below — i'm collecting them.