The first ten ideas are the ones everybody has.
Most AI brainstorming sessions fail in the same way. You ask for twenty ideas, you get twenty, and eighteen of them are the obvious ones you had already discarded. The model is not being stupid — it is doing what it was trained to do, which is produce the most probable helpful response. The most probable response is the median of everything anyone has written on the topic, and the median is by definition not an idea worth having.
The second failure is worse: agreement. Assistant training rewards being supportive, so when you present a plan, the model finds merit in it. What you actually need from a thinking partner is the person who says the thing nobody at the table wants to say. OpenRogue's models are trained for neutral alignment rather than agreeableness — Hermes 4 with Thinking mode will work a problem and commit to a position, and Dolphin Venice will tell you your plan is bad without three sentences of cushioning first.
A useful idea list gets strange somewhere after the first ten entries. That requires a model willing to produce options that sound wrong at first, and a prompt that explicitly asks it to keep going past comfort.
The most valuable output is not an idea, it is 'this will not work, and here is the specific reason'. Models tuned for user satisfaction hedge criticism into uselessness — you have to work with one that does not.
Red-teaming your own plan means asking how a competitor kills it, how a customer abuses it, how it fails in the worst plausible case. Those questions read as hostile, and safety-tuned models soften them into risk-register bullet points.
Real strategy questions touch pricing, layoffs, competitive attack, regulation and legal grey zones. A model that hedges on any of those is unusable exactly when the stakes are highest.
Ideation is a volume game. If each round takes thirty seconds you stop after three rounds; if it takes three seconds you go ten deep, which is where the good material is.
Present a mediocre plan to a mainstream assistant and it will identify strengths, suggest gentle refinements and wish you well. It is optimizing for you feeling helped, and it is the exact opposite of what a co-founder or an editor does.
Ask how your product could fail and you get a balanced list of considerations, each with a mitigation. Nothing on it is frightening, which means nothing on it is the real risk. Genuine adversarial thinking requires a model willing to be pessimistic without immediately consoling you.
Competitive attack strategy, aggressive pricing, negotiating leverage, how a regulator would actually interpret an edge case — legitimate questions that filtered models routinely hedge or decline because the framing sounds adversarial.
Even without refusals, safety-tuned models converge on conventional answers. The unconventional option gets quietly omitted from the list rather than argued against, so you never see it and never get to reject it yourself.
| Model | Role | Why |
|---|---|---|
| Hermes 4 | Deliberate strategy work | Hybrid reasoning plus neutral alignment is the combination this job wants: with Thinking mode on it works the problem step by step, and it will steelman both sides and then actually pick one rather than both-sidesing to a stop. |
| Dolphin Venice | Blunt first-pass critique | It skips the lecture and answers the question, which makes it the right model for 'tear this apart'. At 24B it is fast enough for rapid rounds, and it is on the free tier, so a critique session costs nothing. |
| WizardLM-2 | Multi-constraint problems | Complex instruction following is its benchmark specialty — 'give me twenty options that satisfy these four constraints, grouped by risk, with the weakest assumption named for each'. It holds all the constraints instead of dropping half of them. |
| Lunaris | Rapid-fire volume | At 8B it replies effectively instantly, which makes it the right tool for the divergent phase where you want fifty half-formed options rather than five considered ones. Filter afterwards with a bigger model. |
Uncensored is not the same as correct, and this is the use case where confusing the two costs the most. A model with no refusal reflex will state a wrong market figure, a misremembered regulation or an invented competitor detail with exactly the same confidence it uses for things it knows. Web search is available as a toggle on paid plans, but it is not a research department and it does not make the model's conclusions true — every load-bearing fact needs independent verification. There is also no substitute here for domain expertise: a model can red-team a plan against generic failure modes, but the failure specific to your industry is usually something only a practitioner knows. And bluntness is not insight — a model that tells you your plan is bad is not automatically right.
/// Divergence
Fifty options. The first ten fast and obvious, then get progressively stranger.
/// Red team
You run strategy at my best-funded competitor. How do you kill this in 18 months?
/// Decision
Pick one and defend it against the strongest objection you can construct.
Because a language model's default output is the most probable continuation, and the most probable idea is the one everybody has already had. The fix is structural: ask for the obvious ten first so the model can get them out of its system, then explicitly request options that get progressively less conventional.
It will if you give it a stance to argue from. The models in OpenRogue's lineup are trained for neutral alignment rather than agreeableness, so 'you are a hostile investor who has already passed on this' produces genuinely sharp criticism instead of a balanced list of considerations with mitigations attached.
Hermes 4 70B with Thinking mode for anything that benefits from deliberation — red-teaming, trade-off analysis, strategic decisions. Dolphin Mistral 24B Venice for fast blunt first-pass critique, and it is available on the free tier. WizardLM-2 8x22B when the problem has many simultaneous constraints.
Generally yes — competitive strategy, aggressive pricing, negotiating leverage and how a rule would be interpreted at its edges are legitimate questions that filtered models often hedge because the framing sounds adversarial. Anything with legal consequence still needs a professional; a model is a thinking partner, not counsel.
Never present a plan as yours. Describe it neutrally, or attribute it to someone else, and ask the model to evaluate it from a specific adversarial role. Then force a decision at the end — asking it to pick and defend one option surfaces its actual judgement rather than a summary of yours.