Strategy, HDR and Bracketing Prompts
I spend a lot of time talking to groups of people in companies, trade associations and training cohorts about their use of artificial intelligence. One of the first things I try to get across is how broad the technology is. Even if someone doesn’t use models like ChatGPT or Claude, the chances are that they use AI in parts of their life so commonplace that it barely registers. The example I give is smartphone cameras. In the moment between pressing the shutter button and the photo appearing in an album, on-device machine learning is doing all kinds of work: identifying which parts of the frame are sky, faces or foliage and processing each differently, judging exposure and colour balance, and in low light, assembling a usable image from conditions that would have defeated the compact cameras I grew up with. If my old print albums are anything to go by, the contrast is striking. We’re not only taking vastly more photos than at any point in history; the average quality has gone up.
Thinking aloud, this strikes me as a good metaphor for the use of LLMs in writing. In the same way that my smartphone has made me a competent photographer, an LLM brings a below-par writer somewhere closer to the average. But a talented photographer would find that the auto settings of a smartphone camera reduce the choices available to them and produce blander work. The same concern applies to writers with a genuine style of their own.
This matters even more when it comes to the sort of nuanced thinking and writing that goes into a strategy process. In my book, I write about using LLMs to give a different perspective on a data point or piece of analysis: how might it seem to a customer, an employee, a competitor or a regulator? To be clear, this kind of perspective-taking is not genuine insight. Instead, it gives a probabilistic average of how someone in that position might interpret the information, not how a real person would. But those real perspectives can be expensive, time-consuming and difficult—sometimes impossible or unlawful—to obtain. Thus, used carefully, I’d suggest the synthetic version is still valuable.
Thinking about this in the context of the photography metaphor, I wonder if there’s a further refinement, based on how high dynamic range photography works. No camera, however smart, can capture everything a scene contains in a single exposure. Expose the shot for shadows and you lose detail in the highlights. Optimise for the highlights and the shadows disappear into darkness. Either way, information at the extremes isn’t merely smoothed over: it isn’t captured in the first place. HDR solves this by taking multiple exposures of the same scene for each push of the shutter, each prioritising a different part of the tonal range, and combining them into a composite that keeps detail throughout that range.
Note that this is not taking different perspectives and synthesising them. That would be something closer to what Philip Tetlock calls the dragonfly eye in Superforecasting, where many vantage points are aggregated into one view. The persona and perspective-taking technique I described above moves the metaphorical camera to a different vantage point: you ask what the scene looks like from the regulator’s position, or the competitor’s. Bracketing keeps the camera exactly where it is and varies what is prioritised within the same perspective. The two techniques are complementary, and in the context of the metaphor I’m making for strategy, I wonder if the second is the underused one.
In strategy work and prompting alike, the way we ask a question is quite determinative of the answer we get. Instead of exposures that prioritise light or shade, in strategy work we might ask questions that prioritise an optimistic reading of the data or that concentrate on downside risk. We could take the next quarter as the frame or the next five-year plan period. If we don’t explicitly set our preferences, the LLM will fill them in probabilistically. Ask an LLM a one-off strategy question without specifying the parameters and what you get back is the equivalent of the auto-metered, median exposure: balanced, competent, solid in the middle of the range but lacking detail at the edges.
Where this happens automatically with HDR photography, it requires some conscious effort in prompting and combining LLM outputs. When I press the shutter on my camera app, it brackets for me. The HDR version of writing and strategy exists only if you deliberately prompt for it, the way a photographer bracketing manually takes shots that their light meter would advise them against. The process that the phone camera has automated is precisely the discipline the human LLM user has to supply, though even with some manual effort, generating those extra exposures need not take substantially longer than producing one. My experience so far is that the combination will be richer, with more of the edge cases, than any single automatic pass.
Bracketed prompting widens what you can see of a problem. It doesn’t guarantee a genuinely novel frame, which still requires a human oriented in the right direction. But that seems to me the right division of labour: a human points the camera, the model extends the dynamic range, and the human decides whether the picture is a keeper.