A developer with significant experience building AI workflows has published a provocative take that's resonating across the OpenClaw community: a folder full of instructions is not automatically a tool. The analysis, available in both English and French editions on DEV.to alongside a companion video, challenges prevailing assumptions about how AI agents should be structured.

The Skills vs. Tools Confusion

The core argument centers on a fundamental misclassification that has crept into agent architecture discussions. SKILL.md and AGENTS.md files—documents that outline instructions, prompts, and behavioral guidelines—are increasingly being treated as if they're equivalent to actual functional plugins. But according to the analysis, this equivalence is broken at its foundation. The author spent considerable time building AI workflows before arriving at what they describe as a painfully simple realization: documentation about how to do something isn't the same as actually doing it. A skill file can be brilliant in its guidance—perfectly crafted prompts that tell an agent exactly how to approach a problem—but without underlying functionality, it's still just words on a page.

What Real Plugins Actually Require

The piece argues that genuine plugins require what the author calls 'motors'—the actual executable components that perform tasks. These motors can be APIs, scripts, CLI tools, or any functional code that gets invoked when work needs to happen. Skills, by contrast, are the pedagogical layer: they teach AI agents how to use those motors effectively. The distinction matters enormously for reliability and predictability. When a plugin is just a collection of instructions, its output depends entirely on how well an LLM interprets those instructions in real-time. Add actual motors into the equation, and you get deterministic behavior backed by executable logic that doesn't vary based on model temperature or context drift.

Implications for OpenClaw Architecture

For developers working with OpenClaw's plugin ecosystem, this framing has practical consequences. Plugin authors need to ask themselves whether they're shipping functionality or just guidance. If the answer is 'just guidance,' then what they've built might be better classified as a skill pack rather than a plugin. This doesn't mean skills are worthless—they're essential for helping agents understand context, decide when to use certain tools over others, and handle edge cases gracefully. But conflating the pedagogical layer with the execution layer creates brittle systems that break in production even when they test perfectly.

Key Takeaways

  • Skills (SKILL.md, AGENTS.md) teach AI how to USE tools—they're documentation, not functionality
  • Real plugins need motors: executable code, APIs, or scripts that perform actual work
  • Treating instruction files as equivalent to functional plugins leads to unreliable agent behavior
  • The OpenClaw ecosystem benefits from clear separation between teaching (skills) and doing (plugins with motors)

The Bottom Line

This is the kind of architectural clarity the AI agent space desperately needs. Stop dressing up documentation as software—give your plugins actual motors, and let your skills do what they were always meant to do: teach.