Tools Are Not the Same Thing as Know-How

Anthropic shipped skills for Claude in October 2025 and published the format as an open standard at agentskills.io that December, with SEP-2640 finalizing in September 2026 to give MCP a Skills extension (io.modelcontextprotocol/skills). The writeup argues that a skill grants knowledge rather than access, using progressive disclosure to load SKILL.md instructions only when a request matches, and that MCP's transport layer plus skills' procedural payload are both required for agents to execute real work reliably.

Hand an agent a full sales stack. It can search contacts, send email, create a sequence, update the CRM, pull campaign stats. Every action works. Now ask it to run a campaign in a market it has never seen. It will email a contact another team is already working, send a third follow-up to someone who asked to be removed, and write a generic note to a company that manufactures in-house. None of that is a reasoning failure. Procedure was missing. Anthropic shipped skills for Claude in October 2025 and published the format as an open standard at agentskills.io that December. A skill is a directory with a SKILL.md : name and description in YAML frontmatter, markdown instructions below, optional scripts/ and references/ . The loading model is the clever part, and it is progressive disclosure: SKILL.md body, loaded when a request matches the description That is why you can install a large skill collection without paying for it on every turn. One property matters more than the rest: a skill grants knowledge, not access. The instructions can only direct capabilities the agent already has. Authority stays with the tool and credential layer, which is the right place for it. Capability is not additive. A large unstructured tool surface creates its own failures: selection ambiguity, unnecessary action, poor sequencing, duplicated work, and outcomes you cannot measure because no procedure was followed. The bottleneck is what the agent can decide, not what it can reach. A prompt like "write a good cold email" improves writing and settles nothing else. A mature skill carries the objective, the context it requires, decision rules, tool selection, execution order, business constraints, exception handling, an expected output shape, and a way to evaluate the result. The test is simple: could a competent new hire follow it and get a similar result on a bad day? SEP-2640 became final in September 2026, giving MCP a Skills extension io.modelcontextprotocol/skills . It deliberately defines transport, not payload: the file format belongs to agentskills.io, so there is one definition of a skill and one way to deliver it. Servers expose skills/list and skills/get , files travel as ordinary resources, and every file is bound to a SHA-256 digest. MCP connects. Skills describe. You need both. Models converge and well-documented APIs get wrapped. What does not copy easily is the judgment inside a procedure: workflow order that was corrected rather than assumed, edge cases that only appear in production, and rules shaped by real customer behaviour. The full argument, with the sales-specific skill design: Why AI Agents Need Specialized Skills to Do Real Sales Work https://salesruns.com/blog/261005-ai-agents-sales-skills Originally published on SalesRuns.com https://SalesRuns.com/blog/261005-ai-agents-sales-skills .