What Techniques Make AI Think Like a Strategist?

The model gave me a plan. Bullet points, clean logic, reasonable steps. I read it twice and felt that familiar flatness - the sense that something important was missing, that the output was technically correct and strategically useless. So I changed the prompt. Not the task. The framing. I told the model it was advising a board preparing for an acquisition that might destroy the company's culture, and that the CEO already knew what she should do but needed someone to help her think about what she would do under pressure. The response that came back was different in kind, not degree.

The techniques that make AI think like a strategist are about prompting for tension, not answers. They involve constructing the right cognitive frame before asking any question: giving the model adversarial context, forcing tradeoffs, embedding time pressure, and treating uncertainty as a feature rather than something to resolve. Strategic thinking is the ability to hold contradictory pressures simultaneously and act anyway. The same mental moves that create strategy in humans can be scaffolded into AI responses - if you know which levers to pull.

The Frame Determines the Output

Ask an AI "what should we do about X?" and you'll get a consultant's slide deck. Ask it "what would a competitor do if they wanted us to choose X?" and you'll get a strategist's paranoia.

This distinction tracks directly to research on analogical reasoning. Dedre Gentner (Northwestern University) showed in her foundational 2003 paper in Psychological Science on structural mapping theory that humans generate better strategic insight when forced to compare structurally dissimilar situations rather than similar ones. The same pressure applies to AI. When you ask a model to reason by analogy across domains, you activate different reasoning pathways than when you ask for direct analysis.

In practice, this means starting your prompt by establishing a frame that contains pressure. Military strategy. Ecological competition. A chess endgame where both players have equal material. The specific frame matters less than the fact that it introduces constraint. Unconstrained AI reasoning tends toward comprehensiveness. Strategic thinking, by contrast, is the art of ruthless exclusion.

The historian John Lewis Gaddis, in On Grand Strategy (Yale University Press, 2018), makes an argument that resonates here: great strategists are defined not by their vision but by their ability to feel, at a bodily level, the resistance between what they want and what the world will allow. You can't replicate that in a model. But you can build a prompt that forces the model to articulate that resistance explicitly.

A related lens comes from Roger Martin at the Rotman School of Management, whose work on integrative thinking - documented in The Opposable Mind (Harvard Business Review Press, 2007) - shows that expert strategists hold two opposing models in mind simultaneously and produce a creative resolution rather than choosing one. Prompts that force the model to articulate both sides of a genuine tension before synthesizing them replicate this structure artificially.

Adversarial Prompting as Strategic Scaffolding

One of the most reliable techniques: tell the model to argue against the plan you're about to present. Not to find weaknesses - that produces polite objections. To destroy it. To assume it will fail and work backwards from the failure.

This is pre-mortem analysis, a method developed by psychologist Gary Klein and documented in Sources of Power: How People Make Decisions (MIT Press, 1998). Klein found that teams who imagined a project had already failed generated 30% more valid risk insights than teams who did forward-looking risk analysis. The temporal inversion - starting from failure rather than working toward it - bypasses motivated reasoning.

When applied to AI prompting, this technique produces outputs that feel qualitatively different. The model stops optimizing for plausibility and starts interrogating assumptions. You have to resist the urge to soften it. "What could go wrong?" gets you a risk register. "Assume this failed catastrophically six months from now - what was the first domino?" gets you strategic thinking.

There's an edge case worth naming. Pre-mortem prompting works poorly when the decision has already been made and you're rationalizing rather than deciding. In those cases, the technique produces a kind of theater - the model generates failure scenarios that feel rigorous but that you'll unconsciously discount. If you notice yourself nodding along without updating anything, you're probably past the decision point and running the exercise for comfort.

Forcing Tradeoffs to Eliminate Strategic Fog

Real strategy requires choosing. Most AI outputs avoid choosing. They hedge, qualify, present "multiple perspectives," balance considerations. This is epistemically virtuous and strategically useless.

The technique that breaks this pattern is explicit constraint imposition. Before asking for analysis, tell the model it can only recommend one option and must defend it against the strongest possible objection to that single choice. Alternatively: "You have one sentence to advise the CEO. What do you say?"

This is not about getting a short answer. The compression forces a prioritization that AI normally distributes across paragraphs. Daniel Kahneman's research on the focusing illusion - documented in Thinking, Fast and Slow (Farrar, Straus and Giroux, 2011) - shows that narrowing attention to a single dimension temporarily suppresses the noise from adjacent dimensions. Strategists do this instinctively. Prompts can replicate it artificially.

A related technique is what I've started calling the irreversibility test. Before asking for strategic advice, I prompt the model to first identify which elements of the decision are reversible and which are not - and to reason about the irreversible elements first. Decisions that are hard to undo deserve asymmetric analytical weight. Most AI reasoning treats all factors as equally revisable, which is how you get strategically naive outputs.

Temporal Layering: Short-Term and Long-Term in Friction

Strategy lives in the gap between what's true now and what will be true later. Getting AI to reason across time horizons requires deliberate construction.

Research by Elke Weber at Princeton's Center for Policy Research (2022) showed that expert strategic forecasters differ from novices not in their ability to predict outcomes, but in their ability to hold multiple time horizons simultaneously without collapsing them. They can say "this is the right move for the next 90 days and the wrong move for the next three years" without feeling compelled to resolve the contradiction.

Philip Tetlock's decade-long Good Judgment Project, summarized in Superforecasting: The Art and Science of Prediction (Crown Publishers, 2015), further established that the best forecasters actively resist collapsing temporal uncertainty - they track probabilities across distinct horizons rather than averaging them into a single confident estimate. Applied to AI prompting, this means separating your time frames rather than letting the model blend them.

Most AI prompts collapse time. They ask for a recommendation, implicitly assuming a single horizon. Separating prompts by horizon - explicitly asking for 30-day, 12-month, and 5-year analysis as distinct frames rather than as one unified answer - produces outputs where the tensions between horizons become visible. And those tensions are often exactly where the strategic insight lives.

What happens when the horizons conflict? That's the question you should be asking the model, not a human. Run both analyses, then prompt: "These two recommendations contradict each other. Under what conditions does the short-term logic win? Under what conditions does the long-term logic win?" The model's answer to that conditional question is closer to strategic thinking than either analysis alone.

The Role of Persona and Cognitive Constraint

There's something slightly uncomfortable about this technique, which is probably why it works.

Asking a model to reason as a specific historical strategist - Winston Churchill deliberating over Operation Torch, or Reed Hastings deciding to split Netflix's DVD and streaming businesses - does something that generic role-play doesn't. It attaches strategic reasoning to a documented body of decisions, trade-offs, and failures. The model isn't inventing a strategic personality; it's accessing patterns from a known record.

Research from Ethan Mollick's group at the Wharton School of the University of Pennsylvania - published across several papers between 2023 and 2024 on AI-augmented cognition - suggests that persona constraints produce more differentiated and useful AI outputs for complex cognitive tasks than open-ended prompting. The constraint does the work. The persona is just the container for the constraint.

Use this technique carefully with decisions that require genuine moral judgment. Asking AI to reason "as a competitor who wants to harm your company" is strategically useful. Asking it to reason "as someone with no ethical constraints" is a different thing entirely, and the outputs become unreliable in ways that aren't obvious until later.

Limitations

These techniques work. They also don't transfer as well as they should.

The research on strategic AI prompting is largely anecdotal or based on small studies. Mollick's work is promising; Klein's pre-mortem research applies by analogy but wasn't conducted on AI systems; Tetlock's forecasting research predates current-generation language models. There are no longitudinal studies showing that organizations using adversarial prompting techniques make better strategic decisions over time. That evidence doesn't exist yet.

More importantly, the techniques described here depend on the prompter already having a meaningful level of strategic literacy. If you don't know what a good strategic insight looks like, you can't evaluate whether the model produced one. These methods amplify existing capability - they don't substitute for it. Someone who has never done strategic planning won't become a strategist by learning to write better prompts. That's a harder claim to sit with, but it's accurate.

Finally, there's a genuine risk of over-engineering. Every technique in this article can be applied so deliberately that the thinking becomes mechanical. Strategy requires a kind of play. Sometimes the best prompt is the bad one.

FAQ

Can AI actually do strategic thinking, or just simulate it?

The honest answer is that the distinction matters less than the output quality. If adversarial prompting and temporal layering produce insights that change decisions, the philosophical question of whether the model "really" thought strategically is secondary. Evaluate outputs, not processes.

What's the single highest- technique for someone starting out?

Pre-mortem prompting. Ask the model to assume your plan failed and work backwards from that failure. It's immediately useful, requires no expertise to evaluate, and breaks the default AI pattern of presenting plans as though they're more reliable than they are.

Do these techniques work better with some AI models than others?

Yes, noticeably. More capable models handle adversarial context and persona constraints without collapsing into inconsistency or defaulting to generic caveats. The techniques described here were developed working primarily with frontier models; results vary with smaller or less capable systems.

When should I NOT use strategic prompting techniques?

When you need information retrieval, summarization, or execution support. Strategic framing adds friction. For well-defined tasks with clear right answers, that friction is waste. Reserve these techniques for genuinely ambiguous decisions where the quality of reasoning matters more than speed.

Strategic prompting connects directly to broader questions about how AI changes the cognitive division of labor between humans and machines - explored in depth in The Last Skill. From here, the adjacent territory worth exploring is how to structure AI-assisted scenario planning, and how to recognize when a model is performing confidence rather than reasoning. The difference is subtle and consequential.