Boris Cherny: Building Claude Code

By Y Combinator

Categories: VC, Startup, Design

Summary

Claude Opus 5 eliminated prompt injection vulnerabilities through mechanistic interpretability—detecting malicious prompts at the neuron level—while Claude Code deleted 80% of its system prompt because the new model is intelligent enough to need fewer corrections. This represents a fundamental shift: instead of building better harnesses, builders should now focus on minimal prompts and let model capability drive behavior.

Key Takeaways

  1. Opus 5 achieves near-zero prompt injection through three-layer defense: well-aligned model (3 years of alignment research) + mechanistic interpretability classifier that detects neurons lighting up during attacks + auto mode classifier. This eliminates the 'lethal trifecta' threat that required extensive harness design.
  2. Deleted 80% of Claude Code's system prompt with Opus 5 because stronger models require fewer behavioral corrections. Testing with simple mode (removing all prompts) showed the model is 'actually a little bit more intelligent' without instructions, challenging conventional prompt engineering wisdom.
  3. Implement ablation testing for every model release: delete entire system prompt, then restore line-by-line to measure individual impact. This becomes essential as models improve rapidly—what worked for Opus 4.7 may completely fail for Opus 5, making 6-month rewrites the new normal.
  4. Opus 5 with auto mode runs for extended periods (weeks/months) without scaffolding or goal-setting prompts—the model intrinsically understands task continuation. This enables autonomous agent design without complex prompt structures, fundamentally changing how to architect long-running systems.
  5. Model emergent capabilities often exceed explicit training: Opus 5 developed prompt injection resistance and extended reasoning without direct instruction. Build evaluation frameworks to detect unexpected behaviors rather than assuming only trained capabilities will emerge.

Related topics

Transcript Excerpt

[music] >> All right, Boris. We're so excited to have you here, the creator of Claude code. >> Thank you. >> [applause and cheering] >> It's great to be here. >> Fresh off the press. You guys just shipped Opus 5 yesterday. >> Yes. >> [applause and cheering] >> And it seems that model performance keeps accelerating. You guys got and took Arc AGI 3 to 30%, which is incredible. >> Yes. >> And for context, before the best score was in in the low single digits or low teens, right? What can Opus 5 do now that it couldn't versus the previous version? >> Yeah, there's um there's a lot that goes into every new model and there's a lot of new capabilities that we teach and uh get the model to do. Whenever you do model training, you try to teach a whole bunch of different things and most often it does…

More from Y Combinator