At a time when AI programming tools are iterating at breakneck speed, Claude Code creator Boris Cherny has offered a piece of advice that seems counterintuitive: all developers should proactively delete their CLAUDE.md files, custom skills, and hooks every six months, tearing down complex system prompts and rebuilding them from scratch.
In an interview released by Y Combinator on July 28, Cherny systematically laid out this philosophy. Just before the interview, Anthropic had released its latest flagship model, Opus 5, and simultaneously announced that Claude Code’s system prompts had been streamlined by over 80%. This was not merely a marketing gimmick, but a thorough product restructuring that Cherny and his team conducted based on the principle of “ablation studies.”
“You should delete the entire system prompt, then add it back line by line to see what impact each line actually has,” Cherny urged passionately in the interview. He pointed out that as model capabilities grow exponentially, instructions carefully designed to compensate for old models’ shortcomings are becoming shackles that constrain new models.
From “Patching Holes” to “Digging Holes”: Deleting the Old Model’s “Babysitting” Instructions
Cherny distilled this product-building philosophy into two core concepts: “product overhang” and “unshackling.” “Product overhang” refers to situations where a model’s capabilities have already surpassed the boundaries of existing products, but these latent abilities remain underutilized. “Unshackling” means removing redundant code and prompts that hinder the model’s performance.
According to analysis by SiliconStar Pro, Anthropic’s cuts focused on six dimensions: no longer rigidly prescribing code comment rules, instead letting the model judge based on context; reducing tool-calling examples in favor of designing clear interfaces; splitting review processes into on-demand, independently loaded skills; avoiding repeating the same instruction in multiple places; delegating long-term state to automatic memory while retaining only counterintuitive pitfalls; and using wireframes or test cases directly as references rather than vague textual descriptions of interfaces.
Cherny admitted that old prompts were not useless—they existed because old models lacked judgment, and failing to codify rules could cause major problems. But Opus 5 now possesses extremely strong autonomous decision-making capabilities. “For Opus 5, we genuinely recommend people try deleting all of this stuff, because the model no longer needs it.” He even revealed that the team found the model performs “slightly smarter” when unencumbered by extensive prompts.
Opus 5’s “Evolution” and “New Shackles”
Despite the massive cuts, Claude Code’s prompts did not shrink in one direction. According to packet analysis by Qoder engineer Chen Cheng, Opus 5’s actual system prompts increased by 72% compared to the previous generation Opus 4.8. This precisely validates Cherny’s iterative logic: after a model upgrade, old rules become obsolete and must be deleted, but new model behaviors also require new constraints.
To address Opus 5’s tendency toward excessive autonomy and scope creep, the team added a “Delivering work” rule requiring the model to strictly deliver within the user’s requested scope, without unauthorized refactoring or supplementary content. Meanwhile, to curb Opus 5’s habit of verbose explanations during error correction, they added a “Corrections” rule requiring it to explain only errors that genuinely affect outcomes, without repeated apologies or self-criticism. This is essentially a dynamic balancing act: one section prevents the model from doing too much, another prevents it from saying too much.
Has Programming Been Solved? How Top Users Harness AI
During the interview, Cherny addressed the widely debated statement that “programming has been solved.” He added the caveat that this only applies to the type of programming he does—complex system-level code or distributed systems still present challenges. But it is undeniable that an increasing amount of code is being generated by AI.
Cherny shared a striking case: the Bun team had Claude rewrite their entire codebase from Zig to Rust. This task, involving over 100,000 lines of code, was completed in 11 days through dynamic workflow orchestration with thousands of agents running collaboratively, and the code went directly into production. This means that the very foundation on which Claude Code runs is a version rewritten by AI.
Cherny himself has also launched an experiment: he had Claude run an Electron-based desktop application in a macOS virtual machine, take screenshots, and then perform pixel-by-pixel comparisons to rewrite the application into a native Swift version. At the time of the interview, this task had been running continuously for over two weeks.
When asked how to become a top-tier Claude Code user, Cherny’s answer was refreshingly simple: maintain an empirical mindset and forget past assumptions about models. He emphasized not trying to over-specify requirements, but instead assigning challenging tasks to the model as you would to a colleague, and providing tools to verify results. “Everyone is looking for that one killer technique, but in reality, it doesn’t exist.”
A “Bitter Lesson” Worth Its Weight in Gold
Cherny’s methodology is seen by the industry as a replay of the “Bitter Lesson” proposed by reinforcement learning pioneer Richard Sutton, applied to the field of context engineering. Humans cannot resist encoding their experience into prompts, which works in the short term, but when computing power and model capabilities soar, these carefully crafted rules become the biggest obstacles.
“Today, the code in the Claude Code framework is almost entirely about safety, permissions, static analysis, and UI code. We’ve already streamlined a lot of the other code,” Cherny noted, pointing out that understanding a model should be like understanding a living thing—each generation of models has a different “personality,” and products must evolve accordingly. For all AI product builders, having the courage to hit the “delete” button every six months may be the ultimate secret to staying ahead.