Turning a chatbot into a sparring partner is less about the tool and more about how you ask.
Most designers use AI to produce things: a logo variant, ten headline options, a quick layout mockup. Fewer use it to challenge their thinking before a client does. That second use case just requires a different kind of prompt.
Why "more content" is the default failure mode
Ask a general-purpose model "what do you think of this logo?" and you'll usually get a warm, descriptive summary, not a judgment. That's because the model has nothing to measure the work against. Without stated criteria, an audience, or constraints, it defaults to describing what it sees rather than assessing whether it works. Anthropic's own prompt engineering guidance makes this explicit: give the model a role, context, and success criteria, and the response shifts from commentary to evaluation.
Give the model a point of view, not just a task
Assigning a persona, like "act as a skeptical creative director reviewing this for a pitch," is a documented technique in both Anthropic's and OpenAI's prompting guides. It constrains the model's frame toward judgment instead of neutral description. Pair that persona with a stated objective, "the brand needs to read premium to a board of directors," and you've given it a yardstick. That's really the same discipline behind a good internal design crit, ported into a prompt.
Ask for tension, not consensus
Telling a model to argue against the work produces sharper output than asking it to "give feedback." Try prompts like "list the three reasons a client would reject this" or "steelman the opposing view." These force specificity that vague feedback requests avoid, functioning like an on-demand devil's advocate.
Use constraints as the backbone
Explicit criteria, brand guidelines, a positioning statement, accessibility standards, turn a subjective reaction into a checklist. OpenAI's prompting documentation stresses that stating evaluation criteria upfront reduces generic responses. Asking a model to check a strategy deck against a business objective can surface logical gaps a purely stylistic read would miss.
Where this genuinely helps, and where it doesn't
AI critique is a rehearsal tool: good for stress-testing logic and catching inconsistencies before a real meeting. It has no access to lived audience context, so treat it as practice, not a verdict, and stay transparent with your team about which parts of a review were AI-assisted.
