Glossary · Technique
Adversarial / Red-Team Prompting
Also known as: Red-team, Devil's advocate
Ask the model to attack your idea before defending it. Surfaces the failure modes before they ship.
Try the interactive templateWhen to use it
- Pre-launch audits of products, pitches, plans.
- Decision reviews where confirmation bias is likely.
- Security or risk assessment of any system or process.
- Writing — drafting the strongest counter-position before yours.
When not to use it
- Emotional support contexts — the harsh framing is wrong here.
- Final-pass polish where you want to ship, not re-debate.
- Anywhere the model has already been overly negative.
How it works
- 1Frame the model as a hostile analyst whose only loyalty is to truth.
- 2Forbid politeness and hedging up front.
- 3Ask for the kill shot first — the single most likely reason for failure.
- 4Then enumerate failure-mode categories: slow death, hidden assumption, competitor move, user revolt, regulatory landmine, team failure.
- 5End with the cheapest defense against the most likely failure mode.
Example
Lazy prompt
Review my startup pitch.
Using the technique
Act as a hostile red-team analyst. Find the kill shot, the slow death, the hidden assumption, the competitor move, the user revolt, the regulatory landmine, and the team failure. End with the cheapest defense.
Common pitfalls
- Some models default to balance — 'on the other hand' creeps in. Explicitly forbid it.
- Without specifying categories, the critique becomes generic ('it might not work out').
- The output is meant to be hard to read. Don't soften it before sharing with the team.
Where this came from
Common practice in AI safety / red-team evaluation, popularized by Anthropic's Constitutional AI work.
Related techniques
Try it interactively
The interactive template lets you fill in your scenario and generates a copy-ready prompt that uses this technique.
Open the template