In a recent Y Combinator talk following the release of Opus 5, Boris Cherny—the creator of Claude Code at Anthropic—shared groundbreaking insights into how engineering teams should build with modern AI models. His central argument challenges traditional software engineering and prompt engineering practices: as models become significantly more intelligent, human-written scaffolding, lengthy system prompts, and accumulated instructions become dead weight that actively degrades performance.
1. The Delete Protocol: System Prompts as Scar Tissue
Anthropic deleted 80% of Claude Code’s internal system prompt because Opus 5 rendered most built-in workarounds obsolete. System prompts are read on every turn and incur a tax on intelligence. Cherny recommends deleting your project’s Claude.md, custom skills, and hooks every six months. Rebuild your config line-by-line only when the model repeatedly stumbles on the same task.
2. Exploiting “Product Overhang”
Product overhang is the gap between what an AI model is natively capable of doing and what current software tools allow it to do. Claude Code was born when Anthropic realized models could write full files, yet existing tools only provided autocomplete or read-only chat. Cherny notes that the overhang with current models is larger than ever, offering massive opportunities for developers to build simpler, thin-harness tools.
3. Automated Code Base Rewrites at Scale
To demonstrate model capability, Cherny highlighted how Bun’s 500,000+ line low-level Zig codebase was converted to Rust in 11 days using dynamic multi-agent workflows. The critical enabler was not complex prompt engineering, but a brutal, comprehensive test suite acting as an objective referee to constantly verify the model’s work.
4. Build the Verifier, Not the Prompt
The core skill in modern AI engineering is pairing a task slightly beyond the model’s ability with an automated verification loop. Cherny demonstrated this with a 3-sentence Slack prompt that ran continuously for over two weeks, tasking Claude to rewrite an app in Swift by visually comparing screenshots pixel-by-pixel until complete.
5. Recurring Single-Sentence Maintenance Routines
Instead of manual code maintenance, Anthropic sets one-sentence prompts on timers (cron jobs) to clean up codebases daily. Automated routines like the “abstraction police” independently scan repositories, identify duplicate patterns, delete dead code, and submit pull requests without human intervention.
6. Treat AI as an Empirical Science
Cherny advises developers to discard theoretical computer science priors and static advice from social media. Working with frontier models is purely empirical: give the model a task, observe where it fails, adjust the feedback mechanism, and repeat.
Mentoring question
Looking at your current AI workflows and prompt files, which rules or custom instructions might actually be ‘scar tissue’ compensating for older model limitations rather than adding real value today?
Source: https://youtube.com/watch?v=Egd65CLmb6w&is=iFFZbbW-8zeh3qKg