Four Summaries And A Decision

Most people use AI like a slot machine. You have a task, you describe it, you pull the lever, you get an output. Maybe you iterate a few times, adjusting the prompt. But you're fundamentally asking AI to converge—to give you the answer—and your job is to judge whether it got there.

This is rushing to convergence. And it skips the most valuable part.

The better move is to deliberately diverge before you converge. Not one summary—five summaries. Not one approach—variations along different dimensions. Your job shifts from "evaluate this answer" to "notice what I'm drawn to."

Here's what this looks like. You have a complex internal report. You need an executive summary. The obvious prompt: "Summarize this report." You get a summary. You ask for revisions. You're in slot-machine mode, converging too fast.

Instead: "Give me five different executive summaries of this report."

Now you have options. One is admirably brief but feels thin. One leads with the recommendation—bold, maybe too abrupt. One buries the lead in context. One frames the timeline risk in a sentence you couldn't have written yourself. One overweights a section you realize doesn't matter.

Assemble a collage

You don't just pick the best one. You build a collage from the parts that work. The next prompt might sound like:

"I like how version two leads with the recommendation. But version four has that line about the timeline risk—use that framing but put it up front. And make it shorter, like version one."

You're not evaluating correctness. You're discovering preferences by reacting, then composing something new from elements that resonated. Some of what you're doing is communicating preferences you already had but hadn't articulated. Some of it is discovering preferences you didn't know you had until you saw the options. Both matter. The AI generates raw material. You do the pruning.

You can push further by asking for variation along specific dimensions. Three questions help you choose useful axes: What does it say? (emphasis, what's included vs. cut, level of detail) How is it organized? (narrative arc vs. key findings first vs. implications first) How does it sound? (terse vs. explanatory, confident vs. hedged) Pick axes where you suspect you have an unstated assumption. Each variation surfaces choices you didn't know you were making—you might realize the summary needs to be much shorter than you assumed, or that "implications-first" feels obviously right once you see it.

What everyone misses

You can also do this same diverge/converge move on the problem, not just the solution.

You came to write an executive summary. But what's the real job of this document? Ask the AI to help you diverge on the problem itself: "Give me five ways to complete this sentence: The real job of this executive summary is to..."

You might get:

• "...summarize the report so people can decide whether to read it"
• "...enable a decision without reading the report"
• "...give enough context to ask sharp questions in the meeting"

These aren't variations on a problem. They're different problems. They produce structurally different documents. "Enable a decision" might mean the summary replaces the report for most readers. "Give context for questions" might need provocations rather than conclusions. Most people never diverge on the problem. They arrive with a fixed frame and ask AI to solve within it.

This pattern—diverge on the problem and converge, then diverge on solutions and converge again—has a name. The UK Design Council called it the Double Diamond back in 2005. Importing a product design framework to write a one-page summary might seem like a strange move. But design and product management are disciplines built around specification—figuring out what to make before you make it. That's exactly the work AI shifts onto you. As AI absorbs more execution, expect their tools to show up in unexpected places. The Double Diamond is just an early example.

The real work

That point in the middle—where you've explored framings and commit to one—is a bet. AI can propose possibilities, but you have to prune them to the ones worth pursuing. In the summary example: you decide "enable a decision" is the right frame, but you might be wrong. The report authors might feel their nuance is being flattened, but you're making a judgment call.

And sometimes you'll get it wrong. You'll commit to a problem frame, start generating solutions, and realize halfway through that you framed it wrong. So you loop back to exploring the problem. This isn't failure—it's the process working. You learn what the problem actually is by seeing what the solutions look like.

What you get at the end isn't just a better summary. It's work that feels like yours, because you made meaningful choices that shaped it. You're not pulling a slot machine lever anymore. You're driving.