Everyone Agreed. That's Exactly When You're Most Exposed.
Agreement Is Not Validation
Your AI validated the plan. The team nodded. Nobody raised a hand. That feels like confirmation — but consensus reached quickly is usually consensus that skipped the hard question.
The failure you ship is almost never the one you argued about. It's the one nobody named.
Devil's Advocate exists to manufacture the dissent that didn't show up on its own — deliberately, structurally, before the decision is expensive to reverse.
How It Actually Works
This isn't "be more critical" in a prompt. The upstream LLM gets adversarial tools it has to invoke and leave a record of:
tools = [
start_adversarial_analysis, # Frame the idea, context, and stakes
challenge_assumptions, # Pressure-test the load-bearing beliefs
run_premortem, # Assume it failed — enumerate the causes
attack_tree_analysis, # Decompose how it breaks into failure paths
synthesize_challenges, # Rank the strongest objections to act on
]The output isn't "looks good." It's a prioritized list of how this decision could fail, with the assumptions each failure depends on — the objections you can still do something about.
When to Reach for This
- ✓You're about to commit real time or money to a plan
- ✓The room is unanimous and nobody is pushing back
- ✓A failure would be expensive or hard to reverse
- ✓You need to anticipate objections before stakeholders raise them
Use Cases
Architecture Bets
Before committing to a design, map how it breaks under load, change, and time — while the cost of changing course is still low.
Go / No-Go Calls
Stress-test a launch or migration decision when the downside is severe and the upside already has plenty of advocates.
Breaking Groupthink
When the team aligns too fast, manufacture the dissent nobody volunteered to raise — without making it personal.
Other Reasoning Tools
Different failures need different interventions. Pick the right tool for the moment.
Your AI Agrees. This One Attacks the Plan.
$20/month for pre-mortems, assumption challenges, and attack trees that hunt for the failure you're not looking at.