Force Three Genuinely Different Approaches
Ban the obvious solution and watch what happens. Ask an AI for a plan and it does what a fast, agreeable engine does best: it converges on the first plausible idea and commits. When you ask for "options," you usually get one real proposal flanked by two strawmen — fake breadth that makes the miss feel considered. The fix is to make divergence a hard requirement, not a suggestion.
Why the first idea wins by default
A language model generates the highest-probability continuation given your prompt. If your framing leans toward a solution, the model anchors on it — and anchoring bias is measurable and stubborn in LLMs. Recent studies show strong models stay vulnerable to it even when told to ignore the hint, and simple mitigations (chain-of-thought, "ignore the anchor," reflection) don't reliably remove it. So an unconstrained "give me some options" prompt tends to elaborate the anchor rather than escape it.
Design practice has a name for the antidote: diverge before you converge. The Double Diamond splits idea generation from idea selection on purpose — trying to do both at once is like driving with the brakes on. You want a wide set of genuinely distinct paths first, then evidence-driven narrowing.
The three-approaches prompt
Don't ask for "options." Name the axes so the model can't collapse them into one. A reliable trio:
| Approach | What it forces | The question it answers |
|---|---|---|
| Minimal | Smallest thing that could work | "What if we did almost nothing?" |
| Buy / reuse | An existing library, service, or pattern | "What if we don't build it at all?" |
| Inverted | The obvious approach is banned | "What's the best plan that can't use X?" |
The inverted slot is the one that does the work. Banning the obvious approach removes the anchor and pushes the model into a region of the option space it would otherwise never visit. You're not committing to the inverted plan — you're using it to reveal assumptions the default plan hid.
Make it a rule, not a wish
- State the count and the constraint. "Give me three approaches that differ in architecture, not in wording. One must not use a database." Vague asks get vague breadth.
- Force a differentiator per option. Require each to name its main risk and the one condition under which it's the right choice. If two options share the same risk, they're the same option.
- Reject near-duplicates out loud. "Approaches 2 and 3 are the same plan with different names — replace one." Naming the collapse resets it.
Power move: sample, then compare
Divergence also lives in the decoding. Self-consistency — sampling several independent reasoning paths and comparing where they land — improves reasoning accuracy precisely because different paths surface different answers. You can borrow the idea by hand: run the same planning prompt two or three times in fresh contexts and diff the outputs. Convergence across independent runs is a signal the plan is robust; divergence is a map of the decisions still in play. Either way you learn something the single first answer would have hidden.
The payoff is cheap and large: you widen the space before you narrow it, so you're far less likely to point a week of work in the wrong direction on the strength of an idea the model never really questioned.
Resources
- Double Diamond design process — The Fountain Institute
- Divergent vs. convergent thinking with the UX Double Diamond
- Understanding the Anchoring Effect of LLMs (arXiv 2505.15392)
- Self-Consistency Improves Chain-of-Thought Reasoning (arXiv 2203.11171)
Building an AI feature? Yeda AI designs, audits, and ships production LLM systems.