Systems, not prompts.

Describe what you're testing and it designs a panel of distinct AI personas, has each react in character — gut reaction, what resonates, friction, purchase intent, price perception — then synthesises a verdict and the real-world tests to run next. It's the fast first pass of focus-group research, in minutes.

The judgment: personas are seeded for genuine variety, the panel is allowed to disagree, and the output ends with an honest note that synthetic ≠ ground truth. The skill is in how you brief the test and how you validate it.

Give it a campaign objective and it decomposes the goal into a strategy brief and an asset plan, drafts each asset channel-by-channel, then runs every draft through an automated brand-voice check — flag, fix, rewrite, score. The clearest example of agentic-not-generative: it runs the whole workflow, not one step of it.

The judgment: the brand-voice QA is the guardrail that turns AI speed into consistency at scale rather than inconsistency at scale — and it surfaces what it would keep a human on.

Add competitors with their websites and it researches each one on the live web — positioning, products, pricing, messaging, recent moves — then synthesises a landscape report: per-competitor threat assessment, where the white space is, the threats to watch, and concrete moves framed for your brand. The one agent that acts on the live web rather than recalling from memory.

The judgment: it's a starting point, not a verdict — which competitors matter, what counts as a real threat, and which white space is worth claiming stays a human call.