Stacking roles: the best thing Claude does is disagree with me
Most people use Claude as a generator. Write this, draft that, give me ten ideas. Fine, but it's the shallow end. The highest-leverage thing I do with it isn't generation at all — it's criticism. The model is a tireless, unoffended critic, and if you assign it the right role it will pick apart your thinking in ways your colleagues are often too polite or too busy to.
The catch is that it won't do this on its own. Left to default, Claude agrees with you. It validates whatever you said last — a tendency the field politely calls sycophancy — and for design critique that default is worse than useless, because it feels like agreement from a smart peer when it's just the model matching your framing back to you. You have to explicitly send it into critical mode.
Stacking roles
The technique I lean on is running the same material through different roles in sequence. Each role sees the work through a different lens, and the handoffs are where the value is.
A synthesis I did recently went through three passes. Round one, senior UX researcher: “Synthesize these notes. Flag weak evidence.” That gives me the clean read. Round two, devil's advocate: “Here's my synthesis. Argue against the three main findings. What would a skeptical PM say?” That's where the soft spots show up — the finding I liked because it confirmed what I already believed, the one leaning on a single vocal participant. Round three, product strategist: “Given the synthesis and the challenges above, what are the two most defensible implications for the roadmap?”
The shape is the point: I didn't ask for the answer once. I built it, attacked it, then rebuilt it from what survived. The devil's-advocate pass in the middle is the one people skip, and it's the one that does the work. A finding that survives a deliberate attempt to knock it down is one I can actually defend in a room. A finding that only survives because nobody pushed on it is a liability waiting for a stakeholder to find.
The role you probably haven't tried
One more role worth stealing. Before I run interview questions past a single real participant, I hand them to Claude in this role: “You are a non-technical professional who has never been in a research session before. Read these questions and tell me which ones confused you, which felt like trick questions, and which you'd be tempted to give a good answer to rather than an honest one.”
That last part is the whole point. The gap between the answer people think you want and the answer that's true is where research goes wrong without anyone noticing, and a leading question invites the performed answer. Having the model role-play the confused, slightly eager-to-please respondent surfaces those questions before they cost me a real session. Same trick works on survey drafts: “flag any question that could bias responses and explain why” forces the model out of generation mode and into critique, and asking it to critique its own output is one of the most reliable quality moves I have.
Why the role instruction carries so much weight
A role isn't decoration. It narrows the space of patterns the model draws from. “UX researcher” pulls generic research language. “Behavioral researcher analyzing why a rational person would make this choice” pulls a sharper, more mechanism-focused frame. “Skeptical PM defending a budget” pulls objections you'll actually hear on Monday. A role steers which slice of everything the model has read it draws from, and precise roles pick better slices.
Stacking them multiplies the effect. One role is a viewpoint. Three roles in sequence is a small adversarial process — build, break, rebuild — and adversarial process is exactly what a single designer working alone doesn't have. Most of us don't have a skeptical PM, a behavioral researcher, and a naive first-time user on call to tear at our thinking on demand. The role layer is the closest substitute I've found, and it costs nothing to run.
So: stop asking Claude what it thinks, and start assigning it who to be. A generator gives you more of what you already have. A critic tells you where you're wrong while it's still cheap to fix.