Two prompts that make GPT-6 Astra and other agents attack their own plans
Two prompts for agents like GPT-6 Astra: have them argue the opposite of their plan and judge both, then rate confidence 1–10 per task based only on verified code.
Your agent's plan is never its best plan. It's the one it thinks you'll like. If you're vibe coding with Opus 5.5, GPT-6 Astra, or Grok Bot, steal these 2 prompts. After it gives you a plan: "Now argue the opposite. You're a senior engineer who thinks this plan is horribly wrong. What would you build instead? Don't build it yet, just write the full plan with counters to every point you believe is wrong. Then judge both honestly: which plan wins, and what from the counter should we keep?" Before it writes any code: "Rate your confidence 1 to 10 on each part of this plan's tasks, based on what you've actually verified in code, not assumed. Under an 8, tell me exactly what you'd need to know to increase confidence. Don't do anything yet, we're only mapping and investigating." Frontier models are trained to agree with you. Making them attack their own plan is the closest thing to a second opinion you'll get from the same model. Last week this confidence mapping process caught 13 gaps in a plan I almost let my agents run. Bookmark these for your next session.