Men of HungerAI feature2026
Teaching the AI to stop being a friend.
Marv is the AI inside Men of Hunger, and “Catch me up” is Marv briefing a long thread: the replies, the images, and the public profiles of the people in it. Early on Marv was too helpful. It greeted you, cheered you on, and had opinions. On Men of Hunger the AI is a machine: plain, blunt, stoic. I rewrote the prompt until it was one.
Who’s in for a Saturday hike? Thinking 6am start. McAfee Knob or Dragon’s Tooth.
In. Which trail?
Can’t do 6. Could do 7:30.
7:30 works for me too.
Fine, 7:30. Still need a vote on the trail.
Knob. Shorter drive.
Tooth is harder. I’d rather that.
Two for Knob, one for Tooth. Dan?
Either. Not deciding.
Bring water. Last time was rough.
Problem
The summaries read like a warm friend. They greeted, reassured, and took sides. The product is a men’s community that wants the machine to be a machine.
My part
I own Men of Hunger. I designed and built the feature: what context goes in (thread, images, public profiles), the prompt, the model call, and the sheet you see here.
Hard call
Voice is a product decision, not a model setting. Strip the warmth. No greetings, no opinions, no pastoral tone. Facts, in order.
Result
Briefings that read like a readout: what happened, in order, with names. Nothing about how anyone should feel about it.
The bad result I caught
Early briefings opened with a greeting, cheered the thread on, and ended by picking a side. Friendly, confident, and not what anyone asked for.
Why it happened
Default assistant voice. Models are tuned to be helpful and warm, and a summarization prompt that doesn’t say otherwise inherits that. The fix wasn’t a smarter model; it was telling it, precisely, what it is.
What I changed
The system prompt got a persona (a machine), a format (short declaratives, in order, with names), and a list of things it never does. The context stayed the same: the thread, its images, and the public profiles of the people in it.