Creative Afterburner Ideation
Don't ask AI for "ideas for X." Build a forced-collision prompt with three locked components:
ROLE: You are [narrow, unexpected expert] with deep tacit knowledge in [specific field].
CONTEXT: Evaluate/redesign [specific object/problem], operating under these real constraints: [list 2-3 hard constraints].
ANALOGY BRIDGE: Propose 5 solutions by applying principles from [unrelated discipline] — specifically [name 2-3 concrete mechanisms/principles from that discipline, not just the discipline's name].
OUTPUT FORMAT: For each of the 5 ideas, state: (1) the borrowed principle, (2) how it maps onto the target object, (3) one concrete implementation detail.
Example: "You are a Formula 1 ergonomist. Evaluate a tractor cabin, constrained by all-day 8-hour shifts and dirt/vibration exposure. Propose 5 solutions using heavy-UAV (drone) control principles — specifically haptic feedback, redundant control mapping, and fatigue-adaptive input sensitivity. For each: name the principle, map it to the cabin, give one concrete detail."
The output is raw material for human designers to react to, argue with, and rebuild — not a finished concept.
Progress:
- Step 1: Define the real target object/problem precisely (not "improve X" — "reduce operator fatigue in tractor cabin during 8h shifts")
- Step 2: Pick a role that is narrow and surprising, not generic ("ergonomist" is weak; "F1 pit-crew ergonomist" or "submarine sonar operator UX designer" is strong)
- Step 3: Pick a donor discipline that has zero obvious connection to the target domain, but shares an underlying structural problem (e.g., heavy-UAV control ↔ cabin controls: both are about precision under fatigue/vibration)
- Step 4: Force explicit mechanism transfer — never let the AI just say "apply drone principles," demand it name specific sub-principles
- Step 5: Request a fixed number of outputs (5 is a good default — enough for variety, few enough to stay reviewable)
- Step 6: Run the same object through 2-3 different role/analogy pairs to triangulate — one run is a coincidence, three runs reveal a pattern
- Step 7: Hand raw output to the human creative team as a "provocation sheet," not a deliverable — frame it explicitly as unfinished
Example 1: Input: Need new concepts for a tractor cabin ergonomics project; team thinks AI is useless for creative work. Output: Prompt: "You are a Formula 1 ergonomist. Evaluate a tractor cabin. Propose 5 solutions using heavy-UAV control principles." Result ideas include: (1) redundant control mapping — critical functions (hydraulics lift, PTO) accessible via two independent input paths, borrowed from drone fail-safe redundancy; (2) fatigue-adaptive stick sensitivity — control responsiveness that decreases sensitivity as a vibration/time-on-task sensor detects fatigue, borrowed from UAV turbulence compensation. These are not finished designs — they are starting points the design team argues over and refines.
Example 2: Input: Marketing team dismisses AI-generated packaging ideas as generic. Output: Reframe the ask: "You are a Michelin-starred pastry chef plating desserts (role). Evaluate this shampoo bottle's unboxing experience (context). Propose 5 changes using principles of plate composition and negative space from fine dining (analogy — specify: height variation, deliberate asymmetry, the 'reveal' moment)." The point isn't the literal output; it's that it generates associative material no one on the team would have produced by brainstorming "packaging ideas" directly.
- Always lock three variables explicitly: Role, Context/constraints, Analogy source — a prompt missing any one collapses into generic brainstorming
- Choose analogy sources that share an abstract structural problem with the target, not surface similarity (drones and tractor cabins both deal with sustained operator precision under vibration — that's the real bridge, not "both are vehicles")
- Demand mechanism-level specificity from the donor discipline ("redundant control mapping," not "drone technology")
- Run multiple role/analogy combinations on the same problem before judging quality — the value is in volume and variety, not one perfect output
- Position the output explicitly as raw/unfinished to defuse the "this is just AI plagiarism" objection — the team's job is the synthesis and taste, not the generation
- Keep constraints real and specific (actual shift length, actual budget, actual material limits) — vague context produces vague analogies
- Using a generic role ("a creative designer") — this produces generic output; specificity of role drives specificity of ideas
- Letting the AI pick its own analogy discipline — this defaults to the most obvious, least useful connection
- Treating the output as a final concept to present to clients — it's fuel for the human team, not the deliverable
- Running the prompt only once — one run looks like noise; three runs with different analogies reveal which ideas are robust
- Picking an analogy discipline that's too close to the target (e.g., comparing tractor cabins to car cabins) — no real distance means no real insight
- Skipping the "name the specific mechanism" instruction — without it, the AI will gesture vaguely at the donor field instead of transferring an actual principle