Boris Cherny: Building Claude Code

Video thumbnail: Boris Cherny: Building Claude Code
Jul 27, 202635m 52s video lengthY Combinator

The Signal

Anthropic’s Boris says the latest model, Opus 5, marks a threshold where human-engineered scaffolding often hinders rather than helps. By adopting an aggressively empirical approach—deleting system prompts and harness tools, then adding back only what demonstrably fails—the team has unlocked persistent, long-horizon autonomous tasks that previously required near-constant human oversight.

The Case

The Shift to Empirical Harnessing

  • Anthropic cut over 80% of the Claude Code system prompt for Opus 5, finding that newer models often compensate for old instructions with native capability.3:30
  • The team follows an "ablation first" workflow: they delete existing prompts and skills, test the model, and only reintroduce complexity when repeated failure makes it necessary.5:52
  • Boris frames modern AI product design as "unhobbling" the model, where the primary goal is removing product-imposed constraints to elicit latent performance.10:41

Long-Horizon Agentic Performance

  • Opus 5 enables persistent execution that the team claims can run for weeks or months, a functionality they characterize as unique among current frontier models.1:32
  • A specific Bun codebase rewrite from Zig to Rust ran autonomously for 11 days, utilizing dynamic workflows and local steering to complete the transition.16:03
  • Internal automation is now deeply integrated; the team runs 20 to 30 routines daily—such as dead code removal and test cleanup—across their own app repositories.28:13

Safety and Verification

  • Anthropic asserts it cannot demonstrate prompt injection on Opus 5 when using their three-layer defense, which combines a well-aligned base model with classifiers derived from mechanistic interpretability.2:48
  • Boris emphasizes that successful agentic work relies on task-level verification—such as pixel-by-pixel comparisons—rather than rigid, human-scripted step-by-step instructions.20:29

The 1 Minute Signal Take

This shift acknowledges that as models gain intelligence, traditional "prompt engineering" and rigid system prompts are becoming legacy debts that actively inhibit performance. The takeaway for developers is to treat the AI as an autonomous agent requiring guardrails and verification targets, rather than a deterministic script runner needing constant micromanagement.

Pro Analysis

Why It Matters

This commentary marks the transition from 'AI as a chat interface' to 'AI as a persistent workforce.' The strategic shift from over-prescriptive scaffolding to empirical, long-running agentic loops represents a fundamental change in how software is maintained and scaled, suggesting that the bottleneck is no longer model capability but the human inability to let go of legacy control.

Strategic Implications

Businesses that continue to treat AI as a deterministic, script-following tool will likely fall behind those that embrace autonomous agents. Companies should begin identifying high-volume, low-context maintenance tasks that can be fully offloaded to agentic loops, freeing human talent for high-level architectural decisions.

Evidence & Hype Audit

  • Trustworthiness: Boris provides concrete examples (Bun rewrite, desktop app automation) that bolster the credibility of his performance claims.
  • Bias: The content is heavily slanted toward an internal 'Anthropic-first' perspective and reflects a 'move-fast' engineering culture.
  • Deep Data: While impressive in their specific implementation, these results have not been audited by third parties, and the 'solved' claims remain anecdotal to Anthropic’s specific infrastructure.

Counterarguments

The heavy reliance on Anthropic’s 'three-layer' security stack implies that prompt injection resistance is perhaps less a result of model intrinsic quality and more a testament to aggressive filtering and orchestration, which might not be replicable for independent developers using standard API integrations.

Who Should Care

  • AI Product Managers: To understand the trade-offs between user-facing control/ux and the raw, unhobbled performance of open-ended agents.
  • Engineering Leads: To re-evaluate existing internal maintenance tasks for potential automation.
  • Computer Science Students: To re-center their education on practical problem-solving and cross-domain product skills.

What to do next

  • Audit your existing system prompts; delete the lines that define 'personality' or 'task structure' and perform an A/B test on performance.
  • Implement a verification layer for your most common agentic tasks rather than increasing prompt length.
  • Establish an 'ablation routine' where old tool definitions are removed every time you upgrade to a newer model version.
  • Identify 3 high-frequency maintenance tasks and experiment with a local routine/cron-based agent to handle them.
  • Prioritize learning how to evaluate software quality alongside writing it.
Time saved:32m 25s

Share this

Tags

Written by: 1 Minute Signal Editorial Team