Anthropic ships Skills, and a new shape for distributing AI capabilities lands

On 16 October 2025, Anthropic announced Skills, a packaging format that let teams bundle a prompt, a set of scripts, and supporting resources into a versioned unit that Claude could load into a session at runtime. As Anthropic framed it in the launch post, Skills sat between system prompts (one-off, per-application) and fine-tunes (rigid, expensive, lossy) as a third option for instructing the model on a specific task.

The format itself was a directory with a SKILL.md frontmatter file, a body of natural-language instructions, and zero or more supporting files: scripts the model could execute, schemas it could consult, examples it could read. As reported by The Verge and TechCrunch, Skills landed first in the Claude API as a beta and shortly after in Claude Code. Within a week of launch, the Anthropic-published catalogue of first-party Skills covered tasks from Excel manipulation to PDF generation to PowerPoint authoring.

Why the format mattered

The interesting claim in the launch post was structural. As models got better at long-context understanding and tool use, the distinction between "the model" and "the prompt that turns the model into a particular tool" had been getting blurrier. Skills made the distinction explicit: the model is general; the Skill is specific. The Skill could be versioned, audited, code-reviewed, and shipped through the same processes as the rest of an organisation's software estate. As Latent Space put it in coverage that week, Skills were what fine-tuning had wanted to be in 2023.

Within ninety days of launch, third-party Skills marketplaces had begun to emerge. Claude Code's Skills directory, which surfaced both Anthropic-published and user-created Skills, had crossed several thousand entries by the end of 2025. The format was not adopted verbatim by competing labs but the design pattern, runtime-loadable capability bundles, has clear analogues in subsequent OpenAI and Google product releases.

Skills are what fine-tuning wanted to be in 2023.

The audit story has been the part most discussed inside enterprises. A Skill, as a versioned artifact in source control, can be reviewed and signed off in the same way as any other piece of software. A system prompt buried in an application config file cannot. Compliance teams at regulated organisations have, by 2026, generally taken the position that Skills sit on the right side of the line between model behaviour and software.