Julietta Yaunches

AI engineer & researcher at NVIDIA. (the opinions on this site are mine alone)

Persona Feedback

Andrej Karpathy

Excited By

Concerns

Would Reshape As

“Let’s start with the simplest possible case. You have one feature to build, one test suite to pass, one context window to fill. Here’s exactly what happens at each step…”

I’d want to see a diagram of the actual information flow - prompt → agent → code → tests → feedback → loop or exit. Where does the human intervene? Where are the automated gates?

The piece should include runnable examples. Show a real CLAUDE.md snippet. Show a real acceptance criterion. Show what the agent actually produces. Let the reader implement your workflow after reading.

Missing


Paul Graham

Excited By

Concerns

Would Reshape As

Start with an anomaly: “I kept noticing that my 3-session parallel workflow outperformed colleagues running 15+ coordinated agents. Here’s what I think is happening…”

Then the exploration: why does coordination fail? What makes independence work? Is there something about the nature of LLM context that makes isolation superior?

The insight: maybe Gas Town isn’t the future for most work. Maybe the craftsman’s workshop scales differently than the factory.

The implication: you don’t need to become Stage 7. There might be a different path entirely.

Missing


Martin Fowler

Excited By

Concerns

Would Reshape As

I’d structure this as a pattern description:

Pattern: TDD-Gated Parallel Sessions

Missing


Synthesis: Cross-Persona Patterns

Agreement

All three personas would agree:

  1. Technical specificity is essential - Show the mechanism, not just the philosophy
  2. The observation is valid - Independence outperforming coordination is real and underexplored
  3. The positioning is honest - Admitting “not ready for Gas Town” is authentic

Disagreement

The Key Question

All three would ask, in different ways: “What’s the minimum viable description of this approach that lets someone else adopt it?”

Strongest Angle (Cross-Persona)

The most promising direction combines:

  1. A surprising claim (Graham): “Why I stopped trying to scale to Gas Town”
  2. Technical mechanism (Karpathy): Here’s exactly how TDD-gated parallel sessions work
  3. Pattern formalization (Fowler): When to use this, when not to, and why

The honest “not ready yet” admission could become the thesis rather than an aside: “Most of us aren’t ready for Gas Town - and maybe that’s fine.”