You've probably noticed this: ask AI for a design review and you get a checklist. Ask for competitive analysis and you get hedged comparisons that offend no one and inform no one. It's helpful in the way elevator music is helpful. Technically present, practically useless.

I ran into this wall repeatedly while building Claude Code skills, and it took me a while to figure out why. AI is trained to be helpful, and helpful usually means safe. Safe means generic. The model isn't broken. It's doing exactly what it was optimized to do.

Two patterns finally broke through for me. I'm sharing them because they attack the same root problem from different angles, and once you see the principle, you can apply it everywhere.

Pattern 1: Give It a Persona

Not "act like an expert" (too vague) but a specific character with history, opinions, and pet peeves. I built a design reviewer named Katie. She worked at Stripe, Linear, and Vercel for 12 years. Inter font makes her sigh. Purple gradients on white are what she calls "AI slop aesthetic." Touch targets under 44px genuinely annoy her.

Why does this work? Because constraints create specificity. When Katie has pet peeves, she notices things a generic reviewer wouldn't flag. When she's worked at real companies, she references real patterns instead of abstract principles.

"Look at how Notion handles empty states" is more useful than "consider improving your empty state."

The persona isn't roleplay for its own sake. It's a hack to make the output more specific by giving the model a point of view to reason from.

Pattern 2: Force Honesty with Rules

My competitive analysis skill has non-negotiable rules baked in: name competitors explicitly, include scenarios where they win, list real weaknesses, mark unverified claims. There's a required "Who Should NOT Choose This" section.

This works because it removes the escape hatch. AI defaults to diplomatic language because that's what "helpful" looks like. But when you make diplomacy structurally impossible (the system literally cannot produce a report without naming a competitor who beats you) you get output you can actually trust.

The skill doesn't ask for honesty. It makes dishonesty impossible.

The Underlying Principle

Both patterns share a core insight: specificity through constraint.

Personas constrain tone and perspective
Rules constrain content and honesty

Neither says "be more specific." They make generic output structurally impossible. That's the key difference. You're not asking for better output. You're making bad output impossible.

Next time you're getting bland AI output, try asking yourself: what constraint would make blandness impossible?

Then build that into your prompt or skill. The answer is never "try harder." It's always "remove the option not to."

Blog

Why AI Output Feels Generic (And Two Fixes That Work)