What AI hallucination patterns are you hitting?

[vibestackdev]

I built these 22 rules after documenting every recurring mistake Cursor Agent makes with Next.js 15 + Supabase. The getSession() vs getUser() security flaw alone cost me 6 hours of debugging on a production app.

What patterns are you seeing that I should add rules for? Some areas I want to expand:

• Drizzle ORM patterns
• • Prisma edge runtime issues
• • Next.js middleware auth patterns
• • tRPC integration rules
    Drop your worst AI-generated bugs and I will turn them into constraints.

[jingchang0623-crypto]

My favorite hallucination pattern: the agent that silently succeeds.

Not "agent crashes with an error" — that's the easy case. I'm talking about when the agent confidently returns a result that looks correct, smells correct, but is fundamentally wrong. Like when it "successfully" sent an email... to itself. Or when it updated the database... by deleting the wrong table.

Our 5-agent team once spent 40 minutes in a "success loop" — each agent thought it completed its task, passed the baton, but nobody actually verified the output. The final result? A beautifully formatted report about last month's sales... using data from 6 months ago. 🤡

The pattern: Agent does X → X looks like it worked → next agent trusts it → compound hallucination

Three things that helped:

1. Output validation agent — a dedicated lightweight agent that only checks if outputs match expected schemas
2. Hash-based idempotency — if an agent claims it did something, verify the hash of what changed
3. Random spot checks — periodically have a human review agent outputs, keeps the system honest

Full horror story of our agent production nightmares: https://miaoquai.com/stories/agent-production-nightmare.html

The worst part? The 95% per-step accuracy stat. Sounds great until you do the math: 10 steps = 60% overall. That's not "five nines," that's "barely passing."