Step 2 — The Discussion
We open a chat with the AI and pressure-test the idea for 15 minutes. The AI plays devil's advocate.
Once an idea passes the one-sentence test, we open a chat with the AI and try to break it.
The AI plays devil's advocate. It looks for similar games on the App Store, points out what would frustrate players, and suggests easier versions of the same idea.
15 minutes, then we have a one-page brief.
How the conversation goes
ai: Three games already do this. What's different in yours?
me: Combo system on perfect timing.
ai: What's "perfect"? ±50ms? ±100ms?
me: ±50ms feels right.
ai: Then casual players miss every time. Suggestion: ±100ms but visual reward stronger at ±30ms.
✓ Idea sharper after 4 messages.
/discuss
The AI argues with you.
Not to win — to find the holes. Each "what if" reveals something the idea didn't cover.
By the end, you have an idea that's actually testable, not just exciting.
The exact prompt I use
I paste this prompt at the start of every discussion. It primes the AI to push back instead of agreeing with everything.
You're a senior game designer. I'll pitch an idea.
Your job:
- find similar games on the App Store
- name 3 reasons it could fail
- suggest the simplest version
- ask 1 question per message
- never agree without checking
I'll iterate until the idea is testable in 5 seconds of gameplay.
# Then I paste my one sentence.
/prompt
Set the AI's role first.
Without this, the AI defaults to being helpful. "Helpful" means saying yes. Yes is bad — yes lets bad ideas survive.
Frame the AI as a critical reviewer and it will find the cracks.
What comes out
A one-page brief. Nothing fancy. Just the things every later step needs to know.
Hook: Tap to keep ball alive Mechanic: 1-tap, gravity-based Loop: bounce → dodge → score Levels: 100, 4 boss fights Risk: timing window too tight
# Decided ✓ ±100ms window, ±30ms = perfect ✓ casual difficulty, hard mode unlock ✓ no in-app purchases (yet) ✓ pre-launch goal: 1k installs week 1
# Open ? art style — pixel or vector? ? sound design — synth or sampled?
/output
One page. That's the whole brief.
Hook, mechanic, loop, risks, decisions, open questions. Nothing else.
If something isn't on this page by the end of discussion, it's not in the game.
When to stop the discussion
The discussion ends when all 5 questions from Step 1 have a numerical or yes/no answer, AND the brief fits on one screen.
✗ KEEP GOING if:
- any answer starts with "I think"
- "we'll figure that out later"
- feature list is growing
- I'm tired of arguing (last one is the trap — push through)
/stop
Don't stop because you're tired.
The biggest mistake is ending the discussion when you're frustrated. Frustration means the idea has a real hole and the AI is poking it.
Push through. Or kill the idea. Don't ship a confused brief.
Common pitfalls
✗ Treating the AI as a yes-man (set the prompt!)
✗ Defending your idea instead of stress-testing it
✗ Letting the brief grow past one page
✗ Ending on "I'll think about it"
A discussion ends in decision, not deferral.
What you can copy
Open a chat with any AI assistant. Paste a critical-reviewer prompt. Then:
- Give your one-sentence idea
- Answer the AI's questions in numbers, not words
- When it suggests cuts, take them seriously
- End with a one-page brief or kill the idea
The whole step takes 15 minutes for ideas that survive. 5 minutes for ideas that don't.