Generate Throwaway Concept Directions on Demand

Workflow area: Ideation & Critique · Tier: Foundation · Hands-on time: ~30 minutes

The expensive part of divergence was always production cost — five genuinely different directions used to mean a week of mockups, so teams explored two and anchored on the first. Generation tools collapse that cost to nearly zero. The new risk is the opposite failure: the generated frames look finished, and finished-looking things have gravitational pull. This playbook uses the speed while keeping the frames disposable on purpose.

When to use — and when not to

• Use it at the fuzzy front of a featurewhen the problem is understood but the solution space hasn't been walked yet.
• Use it to break your own fixationwhen every sketch you make is the same sketch, forced divergence is the cure.
• Use it to make critique concretestakeholders critique five artifacts far better than they answer "what do you think we should do?"
• Skip it when the direction is already decidedgenerating alternatives for a decision already made is theater that erodes trust in the process.
• Skip it for polish-stage workrefinement needs your design system and your judgment, not divergence; generated look-alikes pollute late-stage files.

The recipe

1. 1.

   Write a direction-forcing brief. One paragraph: the user, the job, the constraint. Then force divergence along named axes so you don't get five variations of one idea. In Claude or ChatGPT first — text before pixels:

   Copy

   I need 5 genuinely different concept directions for this design
   problem. Diverge on APPROACH, not styling.
   
   Problem: [e.g. "Approvers rubber-stamp expense requests on mobile;
   we need review to be fast AND considered"]
   User & context: [ONE LINE]
   Hard constraints: [e.g. "must work one-handed; no new permissions"]
   
   For each direction give:
   - NAME (two words)
   - CORE MECHANIC — what makes this direction structurally different
   - THE BET — what must be true about users for this to win
   - THE OBVIOUS FAILURE — why this direction might lose
   At least one direction must remove UI rather than add it. At least
   one must challenge an assumption in my problem framing.
   

2. 2.

   Turn the strongest directions into throwaway frames. Feed each direction's name and mechanic into Figma Make or v0 as its own generation — one prompt per direction, explicitly restating the mechanic so the tool doesn't regress to a generic dashboard. For mood and visual-territory exploration rather than UI structure, Midjourney does the same job at the aesthetic layer. Label the output page "THROWAWAY — do not ship" in the file. Literally. The label does real work when a PM screenshots the wall.

3. 3.

   Harvest ideas, not screens. Walk the five directions with your team and extract transferable mechanics into a list. Good harvest notes look like this:

   Copy

   KEEP: "queue triage" mechanic from Direction 2 — batch-approve
   safe items, full review only for flagged ones
   KEEP: the risk-flag summary line from Direction 4
   KILL: Direction 1 (chat-based approval) — bet requires trust that
   doesn't exist yet
   NOTE: Direction 5's "no-UI" framing exposed that half the approvals
   shouldn't need a human at all — feed to PM
   

4. 4.

   Re-design the winner on-system. Open your real component library and design the harvested direction with actual tokens, real components, and your judgment about hierarchy and flow. This is not tracing the generated frame — it's designing the mechanic properly. The generated frame answered "which direction?"; your file answers "how, exactly?"

5. 5.

   Archive the throwaway page. Keep it in an archive section with the harvest notes attached — it's the "we tried A/B/C" evidence that pb-09's decision trail will want. Then confirm nothing from the generated frames leaked into the shipping file: no pasted components, no eyedropped colors.

Make it yours

• Claude for the divergence, Figma Make for the rendering: the two-step split (text directions first, frames second) consistently beats prompting a UI tool cold — the text step is where genuine divergence happens.
• v0 when the concept is interaction-heavy: for directions whose difference is behavior rather than layout, a clickable v0 artifact communicates the mechanic in critique better than a static frame.
• Gemini or ChatGPT as the fixation breaker: paste your own current sketch and ask "generate three directions that reject this one's core assumption" — a cheap adversarial diverge when you're the anchor.

Quality check

• The five directions actually differ in mechanicif they're one idea in five outfits, regenerate with harder divergence axes; styling variety is not divergence.
• Each direction's bet is stateda direction without a falsifiable bet can't be evaluated, only admired.
• The throwaway label is on the pageand everyone in the critique heard the rule said out loud.
• The harvest list exists in writingmechanics extracted, kills reasoned; without it the exercise was a slideshow.
• The shipped design is built from your libraryaudit the final file for look-alike components and off-token values that drifted in from generated frames; invented patterns that resemble your system are the subtlest leak.