Sobot wrote a guide to AI Agent Skills for customer service, and its definition is cleaner than most of what the agent market offers: a Skill is “a reusable method for handling one type of customer issue.” It tells the agent when to take the issue on, how to move it forward, which resources to use, and when to stop or hand off. The post contains no hard numbers, and that is to its credit. What it offers is a division of labor. Knowledge supplies facts, Tools connect systems, Workflows keep fixed steps predictable, and the Skill decides how those three combine for a given issue.
This is product marketing and it ends in a free trial, but the concept holds up apart from the pitch. A product fault is not a question with one answer. An agent has to confirm the product, read the error code, try a fix, know when to stop trying, and gather what a repair ticket needs. Sobot’s insistence that a Skill “isn’t a fixed script” and isn’t “an entire SOP pasted in as-is” is advice that survives outside its product. So does the discipline of stating when a Skill must not trigger. The sharpest sentence is a promise about AI generation: “It doesn’t invent rules for your business.” We cannot test that from outside, and readers should treat it as the claim it is. Skills can also be imported as Markdown files, which makes them portable in format, though the post does not say whether they can be exported again.
The catalog shows the older layer of the company. Sobot has 5 API pages, and the ones touching this story are named for seats and people, not Skills. The Sobot Agent API sits beside Sobot basic-public/service/坐席信息管理, whose Chinese name refers to seat information management, and the Sobot User API. None of the five is named for Skills, Knowledge, Workflows or the Resource Center the post describes. That may simply mean these product features are configured in the console and were never meant as public API. It still leaves a customer who wants to version Skills alongside other code with nothing in the record to build on.
Sobot’s Kin Score is 34.9, thin. Discoverability is 78.6, its one strong facet, against contract quality at 49.0, access clarity at 27.6, developer ergonomics at 28.0, operational transparency at 23.7 and contract governance at 4.5. Agent Readiness is 26.4, agent-aware. Agent skills is lit, along with spec presence, auth clarity, error semantics and openapi examples, so the catalog does register something on this front. Agentic access, MCP server, idempotency and rate limit signals are unlit, though Sobot’s own blog carries a piece on MCP for customer support and the record was last scored on October 17. A Skill, by Sobot’s definition, must know when to stop and where the case ends. The API record has not yet told an outside agent where it may begin.