The current approach to creating skills for AI agents, especially in ecosystems like Claude Code, is being criticized for its inefficiency. Instead of using an agent to write instructions, experts suggest moving toward creating full-fledged software tools.
What Happened
The author points out a systemic error: attempting to force an agent to autonomously write a skill for a task it cannot handle. This turns the skill into a useless block of text-based reasoning instead of a real tool. The correct paradigm involves creating a skill as a structured folder containing instructions, scripts, and documentation, but only after the agent has successfully solved a complex task manually.
Context
In modern AI agent development systems, there is a gap between a model's ability to reason and its ability to act within a specific software environment. A true skill should represent more than just text; it should be a set of functional files, such as .sh scripts and .md files, integrated into the project environment.
Why It Matters for the Industry
The industry is on the verge of a paradigm shift: from pure prompt engineering to agentic tool engineering. This means moving from optimizing text queries to designing runtime environments and standardized toolsets. In the long term, this will lead to the emergence of frameworks for describing skills as software artifacts and marketplaces for distributed skills.
Why It Matters for Users
Developers and users should change their approach to interacting with AI: instead of asking "write instructions on how to do this," it is more effective to provide the agent with ready-made tools and documentation. This will allow the LLM to transform from a text generator into a reliable coordinator of software tools capable of operating in specific environments.
What Is Not Yet Known / Limitations
There is a divergence in engineering and legal focuses: while developers focus on infrastructure, legal aspects raise questions regarding agent liability when using external tools.
Sources
Author
Look at AI, Editorial Team