Claude Skills - SOPs For Agents
Based on Sam Witteveen's video on YouTube. If you like this content, support the original creators by watching, liking and subscribing to their content.
SOP-style packaging is becoming a standard way to improve agent reliability by structuring inputs, steps, tool use, and outputs.
Briefing
Standard operating procedures are becoming a core building block for AI agents—not just a human workflow tool. The shift matters because LLM output quality depends heavily on the context and instructions provided, and SOP-style structures reduce variability by forcing inputs, step-by-step actions, tool use, and end-of-task verification into a repeatable pattern.
Anthropic’s newly announced “Claude skills” is the clearest example of this trend. Skills are packaged folders containing instructions, scripts, and resources that Claude can load only when relevant. Anthropic has also prepared pre-made skills for common office and document tasks such as working with Excel files, creating presentations, and formatting documents. The practical idea is that these packages act like reusable context bundles: instead of re-prompting from scratch, an agent can pull in the right instructions and assets at the moment they’re needed.
Claude skills aren’t limited to the Claude chat interface. They’re available through the API as well, where skills can be added to messages API requests via a dedicated endpoint, alongside Claude’s code execution tool for running tasks that depend on those skill assets. Anthropic describes skills as composable and portable—stackable across workflows and usable across Claude apps, Claude Code, and API-driven agent setups. That “load only when needed” behavior is central: it’s designed to keep the model’s context focused rather than bloated.
A key feature is a “skill creator” capability inside Claude settings. It provides interactive guidance to design a workflow, then generates the folder structure and a “skill MD” file that defines the skill’s name, description, steps, syntax, and tool usage. The transcript draws a parallel to how coding agents rely on custom instruction files (like agents.md-style patterns), but with the twist that the system can help author the SOP package itself.
The broader ecosystem is moving in the same direction. The transcript links Claude skills to Claude code plugins and Google’s Gemini CLI extensions, framing them as packaging mechanisms for agent context—whether through CLI commands, MCP servers, or other tool integrations. It also notes that foundation-model providers are diverging in their APIs to better support background tools and agent execution, including examples like Gemini API integrations (e.g., Google search) and agent configurations that connect skills with MCPs and backend code execution.
Finally, the transcript highlights practical considerations beyond functionality: token conservation best practices, the presence of a public repo of skills (including examples with code for packaging and validation), and licensing differences across skill packages (some Apache 2 licensed, others strictly copyrighted). It ends with a recommendation to try Claude skills directly—such as converting YouTube transcripts into LinkedIn posts—and to consider turning off training on skills if protecting proprietary SOPs is a priority. Overall, the direction is clear: chat platforms are evolving into app-like agent platforms, with “context window marketplaces” emerging around reusable SOP packages.
Cornell Notes
AI agents are increasingly built around SOP-style packages that standardize inputs, tool steps, and outputs to reduce variability. Anthropic’s “Claude skills” packages instructions, scripts, and resources into folders that Claude loads only when relevant, with pre-made skills for tasks like Excel work, presentations, and document formatting. Skills are composable and portable across Claude chat, Claude Code, and the API, where they can be attached to messages requests and used alongside code execution. A built-in “skill creator” helps generate the required folder structure and a “skill MD” file that defines the workflow. The same packaging approach is spreading across other ecosystems via plugins and CLI extensions, pointing toward marketplaces of reusable agent context.
Why do SOP-like structures matter more for agents than for humans?
What exactly is a “Claude skill,” and when does Claude use it?
How do Claude skills work across interfaces and the API?
What does the “skill creator” do?
How does this relate to MCPs, plugins, and Gemini CLI extensions?
What practical issues show up when adopting skills—beyond functionality?
Review Questions
- How does “load only when relevant” change the way skills should be designed compared with always-on prompt stuffing?
- What role does the “skill MD” file play in turning a workflow into an executable SOP package?
- Why might token conservation influence how a skill handles PDFs or other large inputs?
Key Points
- 1
SOP-style packaging is becoming a standard way to improve agent reliability by structuring inputs, steps, tool use, and outputs.
- 2
Claude skills are folders of instructions, scripts, and resources that Claude loads only when relevant to the task.
- 3
Claude skills are usable across Claude chat, Claude Code, and the API, including via messages API requests and code execution.
- 4
A built-in “skill creator” can generate the required folder structure and a “skill MD” file that defines workflow steps and tool syntax.
- 5
Skills are designed to be composable and portable, enabling stacking of multiple skills in one workflow.
- 6
The ecosystem is converging on similar packaging ideas across providers via plugins and CLI extensions, even as APIs diverge.
- 7
Adoption requires attention to token efficiency, licensing terms, and whether skills should be excluded from model training.