Creative Diversity Gym

1. Frame the task

Start by defining the problem and your current cognitive route. This prevents AI from becoming the first frame.

What are you trying to create? What is the main constraint? How important is originality? 3/5

Useful enoughBalancedHighly original

Time available

Audience / context optional

Current state

🎯Clear / regulated

Use the full creativity cycle.

😵Overloaded

Reduce fields and go straight to a small causal probe.

🏃Rushing / fast-brittle

Use anti-cliché friction and no immediate launch decision.

🧱Stuck

Start with deliberately bad ideas, then expand.

2. Human-first idea lock

Full flow

AI perturbation is locked until the human-first idea set is complete.

🔒 Locked.Complete the required human-first fields to unlock the next step.

3. AI perturbation panel

This is not a chatbot. It generates prompts for your LLM of choice, then asks you to paste back the response.

Select 1–3 perturbation modes

Cross-domain sources

Used when Cross-domain is selected. Pick at least one.

Paste AI response

One idea per line works best. The tool will import the lines into the Diversity Board.

AI outputs can look varied while clustering around the same frame. You will check this next.

4. Diversity board

Tagging is the training. These are simple diversity proxies, not semantic embedding measures.

Ideas to tag

Diversity proxies

Redundancy score

—

Frame spread

—

Human-origin ratio

—

AI-shape risk

—

Diversity growth proxy

—

Key question: Does each new idea add a genuinely new region, or just polish an old one?

5. Select and test

Re-enter constraints. Select one idea, then convert it into a small causal probe.

Novelty × usefulness matrix

Click an idea, then click a quadrant. “Develop now” opens the causal probe.

🚀 Develop now

High novelty + high usefulness

💡 Save as inspiration

High novelty + low usefulness

✅ Reliable baseline

Low novelty + high usefulness

🗑️ Drop

Low novelty + low usefulness

Causal probe

Idea to test Hypothesis: this idea works because... Prediction: if I test this, I expect... Smallest safe test: I can try this by... Success signal: I’ll know it works if... Failure signal: I’ll know it fails if... Possible confound: the result could mislead me if... Next update: I’ll revisit this on...

6. Transfer card

The output is not the endpoint. The endpoint is whether you can recover the method later without the same scaffold.

Tomorrow’s no-AI transfer challenge

Without AI, generate 5 new ideas for your task using the same method.

Did you recover the human-first idea categories? Did you avoid the same clichés? Did you add at least one new frame? Did you select using novelty and usefulness? Did you create a small causal probe?