I Deleted All My Claude Skills... And Claude Got Smarter

Summary

Nate Herk explores insights shared by Boris Cherny, the creator of Claude Code at Anthropic, during an interview on the Y Combinator channel. In the interview, Cherny revealed that Anthropic removed over eighty percent of Claude Code's system prompt upon releasing upgraded models. Cherny explained that Claude Code is constantly changing, with prompts, tools, and instructions frequently rewritten or deleted across releases. Much of the original prompt was designed to correct weaknesses in older models, but frontier models handle those behaviors natively without explicit prompting. Cherny advised everyday users and builders to delete their CLAUDE.md files, custom skills, and hooks every six months to observe what new models can accomplish on their own.

Herk explains the concept of product overhang and unhobbling models. Product overhang describes latent capabilities already present in frontier models that existing product interfaces and prompts fail to exploit. Hobbling occurs when rigid scaffolding, excessive constraints, and overly detailed system prompts interfere with a model's underlying reasoning. Herk illustrates this through an analogy of assigning work: while a novice benefits from rigid, step-by-step instructions, an experienced worker is hindered by micromanagement. Imposing granular procedural rules on modern models prevents them from applying their full problem-solving capacity to the task.

To evaluate this approach, Herk conducted an experiment using Claude Code to summarize the thirty-five minute Boris Cherny interview into a resource guide. In the first run, Herk used his existing workspace configuration, including CLAUDE.md and custom skills specifying exact styling, headers, and visual layouts. In the second run, Herk executed the prompt in a fresh session with zero custom skills or configuration files. The output from the unconstrained model lacked visual branding elements like header images, but its substantive content was superior, cleanly organizing the material into timestamped core ideas. Herk concluded that while system prompts should still supply necessary business context and repository structure, task-level execution should remain flexible.

Herk emphasizes Cherny's recommendations regarding task design and verification loops. Instead of micromanaging execution steps, users should assign high-level goals, define clear guardrails, and establish explicit exit criteria. The most critical component of effective agent orchestration is equipping the model with verification mechanisms so it can evaluate and iterate on its own output until it meets the standard. Herk compares working with AI agents to managing human team members, noting that effective managers provide context, set high standards, and enable self-review rather than dictating every action.