Keep this open and run it before you build anything with an AI model. Out of the box the model catches your excitement and starts building; this card makes it push back first, while the idea is still just an idea and cheap to change. It's the pass I run before I let anything get built.
1. Before you start — set the model to critique
- [ ] Tell the model to argue against your idea instead of helping you love it. You want the reasons it fails, not the reasons it looks clever.
- [ ] Make clear that agreement is worthless here. A yes with no critique behind it doesn't earn its place.
- [ ] Ask for the holes before any plan, so the first thing you hear is the objection, not the pitch.
2. The three roles of the council
One critic tends to soften and drift back into agreement, so split the work. Give each role one job and point all three at the same idea.
The Defender — the strongest case for the idea
- [ ] What do you really gain if this works?
- [ ] What are you already doing by hand today that this would take off your plate?
The Hole-Finder — where it breaks
- [ ] Where does it fall apart? Which edge cases did you skip past in the excitement?
- [ ] What happens if you trust it too much and stop checking the result yourself?
The Evidence Analyst — facts versus wishes
- [ ] Which parts do the facts support, and which are just what you're hoping for?
- [ ] How often do those hard cases really come up, in your real data rather than in theory?
3. The rule: every point backed by something real
Every objection and every bit of praise has to rest on something concrete: a fact or a real example from your own situation. Not "this probably won't work," but "this won't work, because…" with the specifics attached; otherwise you get nice generalities, now in three voices.
4. The verdict: build, reshape, or kill
- [ ] Build (Go) — it survived the critique and still stands. Start building.
- [ ] Reshape — the core is sound but the critique found a real hole. Fix it and send the idea back through the council.
- [ ] Kill — the weak points reach the foundation. Drop it and save yourself the weeks.
5. After "Go" — the smallest working version
- [ ] Break the project into pieces instead of handing the model the whole thing at once.
- [ ] Build the smallest version that genuinely works, then grow it from there. The first pass doesn't have to be perfect, just real enough to improve on.
- [ ] Decide up front how the model will prove the result works, so "done" means shown, not claimed. That's its own discipline: make the agent prove it works.