A panel of AI agents
debate your work
to make it great

Top actions

1. Document the maze-generation algorithm explicitly — name it, define what a ‘cell’ is formally, state the invariant you maintain, and show the argument for why the maze is solvable — Knuth — ‘Step 7 is doing enormous work and hiding the entire intellectual content of the maze generator. Show me the argument, not just the claim.’
2. Restructure the write-up to lead with the general principle — ‘AI handles aesthetics, algorithms handle correctness’ — and subordinate the 12-step pipeline to it — Karpathy — the reusable pattern ‘gets one paragraph at the bottom when it deserves to be the thesis upfront’
3. Write the failure post — the three things that nearly killed the project before you found the workaround — Spolsky — ‘That’s the post that gets bookmarked, linked, and referenced for five years. The architecture post is useful. The war story is what gets bookmarked.’
4. Expose and explain the pinch-point elimination step: how are they detected, what’s the criterion, and what breaks if you skip it — Knuth and Karpathy both flagged this — Karpathy called it ‘a classic computational geometry headache’

United front

All four experts agreed on the core architectural decision: AI belongs on the aesthetic layer (shape, silhouette, theme), and deterministic algorithms belong on the correctness layer (graph structure, solvability). This isn’t a compromise — it’s the only principled division of labor when one layer has no correctness constraint and the other is a provable theorem about path connectivity. Knuth, Karpathy, Victor, and Spolsky all called this insight sharp and rarer than it should be.

Clashes

Still split

Karpathy vs Victor on write-up format: Karpathy says restructure the prose so the principle leads; Victor says replace the prose entirely with an interactive explorable where parameters are knobs, not sentences — genuine open question about whether the author’s goal is transferring a mental model (Karpathy) or teaching the system’s behavior through direct manipulation (Victor).

Karpathy vs Spolsky on what makes technical writing last: Karpathy says lead with the generalizable framework — it travels to other domains; Spolsky says lead with the specific failures — the war stories are what developers actually bookmark and learn from.