Delete Your AI Layer? Not So Fast.

Boris Cherny, the visionary creator of Claude Code, recently lobbed a grenade into the AI development community with a provocative directive: "Every 6 months, delete your entire AI layer." This wasn't a mere suggestion; it landed like an existential mandate, immediately sparking a firestorm of debate and concern across developer forums and social media channels globally.

This audacious pronouncement sent a seismic shockwave through the ecosystem, largely misinterpreted as a universal dismissal of hard-won prompt engineering expertise. The immediate, widespread conclusion was that cutting-edge models like Opus 5 had advanced so dramatically they rendered all previous meticulously crafted prompt chains, custom logic, and intricate guardrails utterly obsolete. The community grappled with the notion that their painstaking efforts to guide less capable LLMs were now not just unnecessary, but perhaps even detrimental.

The implication stung deeply, striking at the core of developer identity. It suggested that months, even years, of crafting intricate global rules, specialized skills, and sophisticated hooks—developers' entire bespoke AI layer—were suddenly branded as disposable, worthless baggage. This collective effort, seemingly reduced to irrelevance by the relentless pace of LLM innovation, generated widespread community concern. Many developers questioned the immediate value of their substantial investments in building robust AI systems, fearing their hard-earned intellectual property was now obsolete.

Cherny’s provocative "delete your entire AI layer" mandate from the creator of Claude Code initially sounds like a call for digital nihilism. But a closer look reveals he wasn't advocating for a blind purge of hard-won progress. Instead, Cherny champions a systematic, research-backed evaluation process: ablation.

Ablation, a term borrowed directly from AI research, means starting from a completely blank slate. Imagine stripping away every prompt, every global rule, every custom skill you've painstakingly built. Then, you re-introduce these elements line-by-line, systematically assessing the impact of each addition. This meticulous re-evaluation reveals what truly contributes to your AI’s performance and what has become redundant or even restrictive.

This isn't just an academic exercise; it’s a crucial insight into modern LLM behavior. Newer models, like Opus, thrive on higher-level goals and robust guardrails, not overly specific, step-by-step instructions. Overly prescriptive prompting, once a necessity for less capable models, now often chokes their emergent capabilities. Ablation helps you pinpoint these instructional bottlenecks, empowering you to "let the model cook" with minimal, impactful guidance and unlock its full potential.

Cherny’s systematic ablation, while theoretically sound, presents a staggering practical hurdle for most operations. This process demands deleting an entire system prompt, then meticulously reintroducing every skill, hook, and global rule line by line to measure its individual impact. Such exhaustive evaluation exacts an extreme cost in tokens, developer hours, and engineering resources, making it a prohibitive exercise.

This advice, however, stems from a perspective of near-limited resources. As an Anthropic employee, Cherny operates with a token budget most businesses can only dream of. He casually mentions running agent sessions for "two weeks," even specifying "14 days, 15 days" for a single task. This is a luxury impossible for teams paying per-token or constantly battling rate limits on powerful models like Opus 5, where every interaction carries a real-world cost.

For the vast majority of companies, a full-system ablation every six months remains an expensive, disruptive fantasy, not a viable strategy. While the intent to optimize by stripping away unnecessary prompts is laudable, this level of systematic, token-heavy iteration is simply not feasible for typical enterprise settings. It’s a research-grade luxury, divorced from the financial realities of most development teams leveraging Claude by Anthropic or any other LLM.

Instead of wholesale deletion, a smarter strategy for prompt management emerges from a spectrum of context. We must differentiate between always-on "global rules"—the fundamental system prompts defining your AI’s persona and core directives—and the on-demand "skills" or "subagents" that execute specific, modular tasks. This distinction is crucial for pragmatic optimization.

Aggressive ablation efforts must target these global rules first. They represent the highest token cost, consuming resources with every inference, and carry the greatest risk of over-restricting modern LLMs. Boris Cherny’s insight, despite its initial misinterpretation, holds here: many foundational instructions, essential for older models, now actively hinder powerful new iterations like Opus 5.

Therefore, keep your global rules surgically lean, reviewing them frequently to excise any unnecessary directives. Conversely, treat skills and subagents as more durable assets. These specialized components activate only when needed, incurring costs on demand, not continuously. Reviewing them annually, rather than every six months, saves substantial engineering time and mitigates the "billion-token blind spot" of full-scale ablation. This balanced approach prioritizes efficiency and impact, ensuring your AI layer remains sharp without constant, prohibitively expensive overhauls.

What is the AI 'ablation' method?

Ablation is a process of systematically evaluating an AI system's prompts and rules. It involves deleting the entire system prompt and then re-adding each line or component back individually to test its impact on performance with a new model.

Why did Claude's creator suggest deleting your entire AI layer?

Boris Cherny's point was that modern LLMs are so capable they often don't need the overly specific, step-by-step instructions required by older models. He advocates for ablation to remove unnecessary constraints and allow the model to perform more effectively.

Is it practical to perform a full ablation on your AI system?

For most developers and businesses, a full ablation every six months is impractical. The process is extremely time-consuming and token-heavy, making it very expensive and potentially disruptive to ongoing development.

What's a more pragmatic approach to keeping prompts efficient?

A smarter strategy is to focus ablation efforts on the most frequently used and token-intensive parts of your AI layer, such as global rules and system prompts. On-demand context, like skills or subagents, can be reviewed less frequently.