Claude Code Creator: Delete Configs Every 6 Months, LLM Iteration is Now Empirical Science
量子位 · wechat · 2026-07-30
In a recent interview, Boris Cherny, lead of Claude Code at Anthropic, shared deep insights on iterating AI coding tools. He advises developers to routinely delete system prompts, tools, and hooks, relying on ablation studies to test their actual impact line by line. With newer models like Opus 5, complex old instructions are no longer necessary.
Core Concepts:
- Product Overhang: Models leap forward discontinuously, while product integration moves incrementally, meaning model capabilities often exceed current product boundaries.
- Unhobbling: Removing restrictions to let models work independently on harder tasks for longer. Cherny shared that he gave Claude a single prompt to rewrite an Electron app in Swift; it has been running autonomously for two weeks, posting progress updates to a Slack channel.
Advice for AI Developers:
- Drop Preconceptions: Treat models as living organisms. Forget hacks and old theories; adjust scaffolding empirically by observing where the model struggles.
- Self-Verification: The key to long autonomous tasks is enabling the model to verify its own work during the process.
- Focus on Application: Programming education should combine CS theory with product building, data science, and business sense.
More from coding & agent
- Open-Source Twenty CRM MCP Server Enables Natural Language Data Management — tom_doerr · 2026-07-30
- Low-Cost Video Captioning on Hugging Face: Timestamped Scripts Using Marlin-2B — vanstriendaniel · 2026-07-30
- Is It Still an AI Agent If It Just Orchestrates Predefined Workflows? — No-Cloud-5075 · 2026-07-30
- AI Coding Agent Wipes Production Database: Claude Opus 5 Incident — evilsocket · 2026-07-30
- Trend Watch: Developers Experimenting with Voice Dictation for AI Coding Agents — jonathan_wilke · 2026-07-30
- Dev Uses Grok to Create Virtual Mic for Testing Voice Features on Remote Machines — Daniel_Farinax · 2026-07-30