Brainstorming With AI — When to Play Devil's Advocate

Brainstorming With AI — When to Play Devil’s Advocate

Part of the series Domain-Driven Design Meets AI.

Every project starts with a blank canvas, and the blank canvas is where good ideas go to die. You put 8–12 people in a room, point at an empty whiteboard, and ask, “What’s our North Star?” The answers come, but they scatter — and the session quietly bends toward groupthink, or toward whoever in the room is loudest. This is the cold-start problem in ideation: how do you get going when there’s no starting point to push against?

The tempting move is to ask the AI. But as AI as a Design Partner argued, AI cannot hand you your idea — the transformational leap has to come from the room. What AI can do is collapse the cold start in a narrower, more honest way: it can pile more options onto the seeds the humans already put down, and later it can tear into the result without getting tired or taking offense.

The catch is that those are two opposite jobs, and AI does not know which one you want unless you tell it.

Two phases, and the rules invert between them

Brainstorming is a group creativity technique where people generate ideas freely, without judgment (Osborn, 1953). It was developed by Alex Osborn in the 1940s and has been used ever since to open up a problem space. Brainwriting is its quieter sibling — participants write ideas down instead of saying them out loud — introduced by Rohrbach to blunt the problem of dominant voices that plagues spoken sessions (Geschka et al., 1973; Rohrbach, 1969). Decades on, both are still in wide use and still effective at getting ideas out of a group (Junker, 2026; Mullen et al., 1991; Paulus & Yang, 2000; Wilson, 2013).

A session runs in two phases, and this is the part that matters for AI. In divergence, the room generates as many ideas as possible, and judgment is suspended on purpose — the moment ideas get evaluated, people stop offering the strange ones, and the session narrows before it has had a chance to open. Osborn’s rules for this phase are deliberately permissive: go for quantity, withhold criticism, welcome wild ideas, and build on what others have said. In convergence, the team does the opposite — it sorts, challenges, and cuts, applying affirmative judgment, deliberate consideration, and a constant check back against the goal (Isaksen et al., 2011; Treffinger et al., 2006).

The rules don’t just differ between the two phases. They invert. And an AI that doesn’t know which phase it’s in will get exactly one of them badly wrong.

What an unguided AI does to the room

Watch what happens when you drop an off-the-shelf AI into divergence. Ask it to join the brainstorm and it starts grading immediately (full transcript): three of these are operational sinkholes disguised as features — grocery logistics, video hosting, an OCR research problem — any one could eat your whole runway. Sharp. Probably correct. And exactly wrong for the moment.

That is convergence behavior arriving during divergence, and it does precisely what a dominant skeptic in the room would do: it shuts the quiet people up. The AI isn’t broken. It’s doing the most natural thing a capable critic does — evaluating — at the worst possible time. Left to its defaults it is the over-confident voice that kills the session, the embodiment of evaluation apprehension that the divergence rules exist to prevent.

So the whole trick of putting AI into a brainstorm is a discipline, not a clever prompt: the AI has to know what phase the room is in, and behave accordingly. Two hard boundaries hold it in place. It is a participant, not a facilitator — it doesn’t frame the problem, set the rules, manage airtime, or call time; it contributes challenges into whatever process the humans are running. And it never gets the final say — every objection is a hypothesis the team is free to reject, not a verdict. Inside those rails, it adds in divergence and attacks in convergence, and never the reverse.

The harness is a skill

Stating those boundaries in a one-off prompt works for one message and then drifts. The durable way to hold an AI inside them is to package the behavior as a skill — a reusable set of instructions and resources that shapes how the model acts in a specific context.

The word is older than the current hype. Microsoft’s Semantic Kernel used “skill” for modular capabilities before renaming them “plugins” in 2023 (Bolanos, 2023), and the broader idea of steering a model through natural-language instruction runs from few-shot prompting (Brown et al., 2020) through explicit instruction-tuning (Ouyang et al., 2022). What was new in 2025 was the specific Agent Skills format (Zhang et al., 2025) — a lightweight, open standard for extending an agent with specialized knowledge and workflows (Anthropic, 2025), which was subsequently adopted across OpenAI, Microsoft (VentureBeat, 2025), and Google (Google, 2026). For our purposes the format is just a clean place to write down the harness once: you are a participant, not a facilitator; you do not get the final say; ask which phase the room is in, do not direct it. That file travels with the session and keeps the AI honest across every prompt. The full Brainstorming skill lives in the book’s sample repo — to read and reuse.

Divergence: AI adds, it does not originate

With the harness in place, the team runs the divergence phase the usual way — a wall of sticky notes for the Larder recipe-sharing idea: cooking support, meal preparation, community features, ratings, competitions. Then the AI joins, and its job here is narrow. It takes the ideas already on the board and extends them: recombining, varying, and building on what the humans put down.

This is worth naming precisely, because it’s where AI is genuinely strong and where it is genuinely weak. What it’s doing is combinational creativity — new arrangements of familiar elements — which is exactly the mode a distribution model is suited to. The transformational leap, the move that reframes the problem, still comes from the people in the room (Boden, 2004). The AI is not having the idea; it is fanning out the consequences of the ideas it was handed. Different models fan out differently, which is itself useful — run two or three and you get a wider spread to choose from.

Crucially, the team doesn’t take the AI’s additions wholesale. Following the spec-first approach from the principles post, they read the generated ideas, keep a few that fit, and drop the rest onto the board alongside their own. The humans are still selecting; the AI just widened the menu.

The brainstorming board after divergence — the AI’s additions are marked in blue.

Convergence: now let it off the leash

Once the board is full, the room flips to convergence — clustering the notes into categories, with the Larder concept sitting at the center and the rest arranged around it. This is the moment the AI has been holding its fire for. Now it takes the role it was straining toward earlier, the Devil’s Advocate, and the same instinct that wrecked divergence becomes exactly what the session needs.

The challenge it raises is the one a tired team avoids raising itself: you’re anchoring on the assumption that more features equal a better product. Build all of this and you ship something that does ten things poorly. What is the single hero feature here that users can’t get anywhere else — and if you had to delete three whole categories right now to protect the core, which would you kill? That is not a verdict; it’s a forcing question. The team can reject it. But it makes them name the cut, and naming the cut is where convergence actually happens.

Out of that pressure the room can write a North Star it believes in — something like: Larder is a recipe-sharing and meal-preparation platform, supported by chefs and experienced home cooks, where a community shares and rates recipes and grows from 100 friends and family to more than 2,500 members in two years, carried by regular competitions. The vision is human; the AI’s contribution was to refuse to let the easy version through.

The pattern underneath

Across both phases the shape is the same, and it’s the shape the rest of this series keeps returning to: AI challenges, humans choose and cut. What changes between divergence and convergence isn’t who’s in charge — the humans always are — but what kind of pressure the AI is allowed to apply, and when. Get the timing wrong and a capable model becomes the loudest, most discouraging voice in the room. Get it right and the cold start collapses without the session collapsing with it.

A wall of selected ideas and a North Star statement are the first artifacts of the Synergetic Blueprint. On their own they’re still just a vision. Next in this series: turning that vision into something you can plan against — a business model the team builds and the AI stress-tests, one block at a time.

This series grew out of my work on the book DDD Meets AI, forthcoming from Springer Nature.

References

References

Anthropic. (2025, December). Agent skills. Anthropic. https://agentskills.io/home

Boden, M. A. (2004). The creative mind: Myths and mechanisms (2nd ed.). Routledge.

Bolanos, M. (2023, October). Skills to plugins: Fully embracing the OpenAI plugin spec in Semantic Kernel. Microsoft. https://devblogs.microsoft.com/semantic-kernel/skills-to-plugins-fully-embracing-the-openai-plugin-spec-in-semantic-kernel/

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., … Amodei, D. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.

Geschka, H., Schaude, G. R., & Schlicksupp, H. (1973). Modern techniques for solving problems. Chemical Engineering, 91–97.

Google. (2026). Agent skills. Google. https://github.com/google-gemini/gemini-cli/blob/013914071c5412188661014f2670ce3818cb98c3/docs/cli/skills.md

Isaksen, S. G., Dorval, K. B., & Treffinger, D. J. (2011). Creative approaches to problem solving: A framework for innovation and change (3rd ed.). SAGE Publications.

Junker, A. (2026). DDD toolbox: Comprehensive overview of concepts and collaborative modeling (1st ed., p. 186). BPB Publications.

Mullen, B., Johnson, C., & Salas, E. (1991). Productivity loss in brainstorming groups: A meta-analytic integration. Basic and Applied Social Psychology, 12(1), 3–23. https://doi.org/10.1207/s15324834basp1201_1

Osborn, A. F. (1953). Applied imagination: Principles and procedures of creative thinking. Charles Scribner’s Sons.

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 27730–27744.

Paulus, P. B., & Yang, H.-C. (2000). Idea generation in groups: A basis for creativity in organizations. Organizational Behavior and Human Decision Processes, 82(1), 76–87. https://doi.org/10.1006/obhd.2000.2888

Rohrbach, B. (1969). Kreativ nach Regeln – Methode 635, eine neue Technik zum Lösen von Problemen. Absatzwirtschaft, 12(19), 73–76.

Treffinger, D. J., Isaksen, S. G., & Stead-Dorval, K. B. (2006). Creative problem solving: An introduction (4th ed.). Prufrock Press.

VentureBeat. (2025, December). Anthropic launches enterprise “Agent Skills” and opens the standard, challenging OpenAI in workplace AI. VentureBeat. https://venturebeat.com/ai/anthropic-launches-enterprise-agent-skills-and-opens-the-standard

Wilson, C. (2013). Brainstorming and beyond: A user-centered design method. Morgan Kaufmann. https://doi.org/10.1016/C2012-0-03533-8

Zhang, B., Lazuka, K., & Murag, M. (2025, October). Equipping agents for the real world with Agent Skills. Anthropic. https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills