Stop Building AI That Agrees With You

I'm not gonna tell you what to do. I'm just gonna tell you why everything you're doing is wrong
— Luci, Disenchantment

Tell me — when was the last time you ran your incident report past someone whose job was to disagree with you? Not a rubber-stamp review. Not "LGTM" in Slack. An actual, adversarial read from someone who doesn't care about your feelings and has every incentive to find the hole you missed.

Yeah. Never. Because that person doesn't exist at night when you're wrapping up the postmortem. So you ship it, feel good about it, and two days later someone asks one question in a thread and the whole thing collapses.

I got tired of this, so I built a Claude Code skill whose entire purpose is to argue with me.

I named it Luci — after the little demon from Disenchantment who sits on the main character's shoulder, drinks her beer, and tells her the thing she didn't want to hear. You feed it your analysis, your diagnosis, your design doc, etc. And it runs a structured challenge protocol before telling you where you're full of shit.

Not vaguely. Specifically, with alternatives, verification steps, and a confidence rating.

Everyone knows about confirmation bias. What nobody talks about though, is how confident you feel right before it bites you. You dig into an incident, find a root cause that fits, and your brain just... stops looking. Every piece of evidence after that gets sorted into "supports my conclusion" or "probably irrelevant." You don't even notice that you're doing it. That's the whole problem.

The implementation version is worse. You pick up a ticket (or hand it to an AI agent — requirements look clear. It gets built, shipped, and it works exactly as specc'd. Then in production, you discover the ticket never accounted for the case where users have multiple accounts, or the upstream data arrives out of order, or the field that's "always populated" is null 12% of the time.

The ticket was wrong.

The PM wrote it from the happy path. Nobody challenged the edges before it hit the backlog, and an agent won't push back on a bad spec. It'll implement it flawlessly.

The Catholic Church figured this out in 1587. They created the advocatus diaboli — the devil's advocate — whose job was to argue against sainthood candidates. Not because they hated saints, but because the decision was too important to leave to one perspective. Smart people with good intentions still run the risk of talking themselves into incorrect conclusions. That's not a bug, it's the default.