The idea behind Claude Skill is not particularly novel. Cursor started promoting .cursorrules back in 2024, iterating throughout the year until it evolved into MDC rules by January 2025—a mechanism strikingly similar to today’s Claude Skill in both form and purpose. Meanwhile, Claude had just launched MCP around the same time.
On the surface, both Cursor rules and Claude Skill share identical goals: enabling models to consistently follow a set of reusable behavioral constraints within specific contexts. Put plainly, both are about prompt reusability.
So why did Claude turn this concept into a widely adopted, industry-defining capability standard while Cursor did not? The answer lies not in who thought of it first, but in positioning, capability boundaries, ecosystem incentives, and industry structure.
The Fundamental Limitation of Downstream Applications
Cursor is essentially an AI IDE that integrates mainstream LLM capabilities. As a downstream application, Cursor’s control over large language models is limited to the application layer—influencing model behavior through prompt concatenation, system prompts, or runtime conventions. For Cursor, rules remain “external constraints” rather than first-class capability structures within the model’s reasoning process. This fundamental limitation explains why, despite its initial appeal, Cursor rules struggled to evolve into a unified, stable, and transferable paradigm.
As a downstream product, Cursor must interface with virtually all mainstream LLM providers. Different LLMs vary dramatically in how they understand and comply with rules, causing the same rule set to perform inconsistently across models. This capability misalignment makes it nearly impossible for the community to develop truly universal and stable best practices—something painfully evident in the troubleshooting and experience-sharing discussions around MDC rules (see the struggles documented in Cursor’s community post My Best Practices for MDC rules and troubleshooting).
Does Cursor have its own model capabilities? Yes—it has a lightweight self-developed model called Cursor small, primarily used for quick responses and basic completions. However, this model’s capability ceiling falls significantly short of Claude and GPT-class models. Its design prioritizes reducing latency and cost for features like Cursor Tab, not complex reasoning or Agent-level capabilities. This inherent model limitation constrains how deep Cursor rules can actually go.
The Upstream Advantage: Why Anthropic Succeeded
When Claude Skill launched, some in the community dismissed it as “just another form of Cursor rules.” But this overlooks a fundamental difference: Anthropic is an upstream model provider.
This distinction—a model vendor directly supporting capabilities from the foundation—creates several advantages that Cursor simply cannot match:
- Pre-training leverage: Anthropic can integrate Skill capabilities during training and alignment phases, utilizing data and optimization across the entire pipeline.
- Unified ecosystem standards: By modularizing reusable capabilities, the entire ecosystem can share and benefit from them.
- High-capability LLM support: Unlike lightweight models such as Cursor small, Claude offers superior reasoning and long-term state management.
Because Skill is backed by an upstream LLM provider, Claude can build top-down support into model training, alignment, and runtime mechanisms. This deep integration makes Skill a “first-class citizen” of model capabilities, not merely an exercise in prompt engineering. The practical difference is significant: Skills can not only control behavior but also influence reasoning paths, invoke tools, and manage state—becoming truly reusable capability modules.
This follows a familiar industry pattern: when a smaller player pioneers an interesting feature in a niche area, a larger player with stronger infrastructure, more data, and deeper integration often productizes and standardizes it for the entire industry. Claude Skill’s rapid emergence as an industry focal point exemplifies this trajectory perfectly.
Hidden Factors: Ecosystem Incentives and Timing
Two additional factors often go overlooked:
Ecosystem Incentives
As a downstream LLM application, Cursor lacks both the motivation and the appropriate position to serve as a cross-product, cross-model standard-setter. Interestingly, given its investment relationships (OpenAI and Google/Gemini have invested in Cursor, and Cursor is Claude’s largest API customer), Cursor theoretically had the opportunity to elevate Cursor Rules to Claude Skill’s current status. But Cursor’s positioning as an IDE prevented this.
At the time, GitHub Copilot, Codex, Windsurf, and Claude Code were all Cursor’s competitors. Cursor’s primary revenue comes from subscriptions—it wanted users to use Rules within Cursor’s IDE, not migrate to competing AI coding products. Claude Code operates on entirely different incentives: Anthropic’s deeper motivation is to turn every machine into an Agent experimentation container, helping train models and Agents—something hinted at in Claude Code’s source code through its A/B testing components. Combined with MCP’s launch, Anthropic clearly aims to establish capability and interaction standards for the Agent era. They want to be industry leaders. Cursor, by contrast, feels more like a company riding the wave to capture short-term gains (somewhat reminiscent of Docker vs. Google’s Kubernetes dynamic).
Timing
When Cursor rules emerged, the industry’s attention remained focused on Copilot and code completion optimization. By the time Skill launched, Agent-based workflows, long-running tasks, and capability reuse had become explicit demands. The same underlying concept, proposed at different moments, naturally yields vastly different impact.
Beyond This Case: A Glimpse into LLM’s Next Chapter
From the moment Claude began hardcoding these capabilities into the model layer and abstracting them into reusable, promotable standards, we can already glimpse the next phase of LLM development.
Over the past few years, industry discussions have largely centered on whether the Scaling Law is hitting its ceiling: high-quality internet text data is being rapidly consumed, and the marginal returns from simply scaling parameters and data are diminishing. But this doesn’t mean large models are approaching their capability limits. What’s actually hitting a bottleneck is just one narrow form of capability—text prediction.
As LLMs are increasingly deployed for real-world tool invocation, code collaboration, task execution, and multi-step decision-making, we’re beginning to discover stable, reproducible capability patterns through practice. Once these patterns are validated across enough real-world scenarios, their generality and effectiveness become an extremely valuable data asset—and model vendors are uniquely positioned to systematically acquire and evaluate this data.
Under this premise, continuously “hardcoding” mature capability patterns into the model layer is no longer just an extension of prompt engineering—it represents a capability evolution path. Models are no longer merely predicting the most likely next text; they are learning to predict the most reasonable next action.
As this “behavior prediction” capability becomes increasingly internalized within models themselves, the core of LLM competition will shift accordingly—from who can generate the most human-like text to who can most reliably predict and execute the correct next action sequence. Ultimately, the models that can consistently make high-quality next-step decisions in complex tasks will win this race.
From this perspective, Anthropic is leveraging standard-setting to acquire more structured capability data, then using that data to train even stronger foundation models. This isn’t a path with immediately visible returns, but in the positive feedback loop of “capability → data → model → stronger capability,” Anthropic is clearly leading the pack.