Detailed Analysis
A recent discussion sparked by comments from Boris (a creator of Claude Code at Anthropic) has surfaced a counterintuitive insight about how to get the best performance from Claude Opus 4.5, formerly referred to in casual conversation as "Opus 5": more scaffolding, more system prompts, and more custom skills can actually degrade the model's output rather than improve it. According to Anthropic's own engineering team, Claude Code's system prompt has been trimmed by over 80% in recent iterations, not because the product needs less guidance in general, but because newer, more capable models require dramatically less hand-holding than their predecessors. Behaviors that once needed to be explicitly corrected through verbose prompting are now handled natively by the model's improved reasoning, meaning that legacy instructions originally written for older models can actively interfere with a more intelligent system's ability to perform.
This matters because it upends a common assumption among AI practitioners and power users: that heavily customized system prompts, skills, and hooks are inherently additive to model performance. The advice from Anthropic to periodically "delete your prompt, delete your skills, delete your hooks" and observe how the base model performs unassisted reflects a broader philosophy sometimes called "unhobbling"—the idea that as models become more capable, artificial constraints originally designed to compensate for weaknesses become unnecessary friction. The article's author illustrates this with an analogy: giving a 10-year-old highly specific, rigid instructions makes sense because they lack judgment, but giving those same rigid instructions to a seasoned expert with a decade of experience actively prevents them from applying their own expertise. As Claude models mature, treating them like the rigid instruction-follower model becomes counterproductive, and shifting toward lighter-touch guidance—preserving only essential contextual information (file locations, business context, documentation structure) while removing task-level micromanagement—appears to yield better results.
The practical implication for developers and technical users is significant: workflows, prompt libraries, and custom skill sets built around older Claude models (such as Claude 3.5 or earlier Opus versions) may need periodic re-evaluation rather than being treated as permanent infrastructure. Anthropic's internal practice of rewriting and pruning Claude Code's system prompt with every new model release signals that the company views prompt engineering as a moving target tied tightly to each model's specific capabilities and quirks, not a one-size-fits-all discipline. This also suggests a shift in how the developer community should think about maintaining AI tooling—treating skills, system prompts, and automation layers as things to be audited and stress-tested against new releases rather than assumed to be model-agnostic.
More broadly, this development reflects an important trend in frontier AI: as models rapidly increase in raw capability, the value proposition of extensive prompt engineering and workflow scaffolding is shifting. Where earlier generations of language models required significant "hobbling" corrections to behave reliably, newer models increasingly internalize good judgment, reducing the need for exhaustive external guardrails. This has implications beyond just Claude Code, hinting at a future where competitive advantage among AI products may depend less on prompt engineering sophistication and more on knowing when to strip complexity away and trust the underlying model. It also underscores Anthropic's rapid iteration cycle and internal acknowledgment that the tools built around Claude must evolve as quickly as the models themselves.
Read original article →