---
title: "The Folder Travels. The Judgment Does Not."
paper: "The Daily Model"
edition: 2026-10-26
newsroom: Anthropic
edition_id: 2026-10-25T04-15-55-178Z-93a72f61c6-approved
content_hash: d0a46d0b57148936d648d9bad220073ceb9c33a864e576833bae8f9ca3791b62
sources:
  - Donating Agent Skills to the Agentic AI Foundation (https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation) 2026-10-22
  - Building Great Agent Skills: The Missing Manual (https://www.youtube.com/watch?v=UNzCG3lw6O0) 2026-06-29
---

# The Folder Travels. The Judgment Does Not.

A folder is an easy thing to pass from one person to another. Knowing whether its contents deserve to be trusted is harder.

That distinction sits beneath [Anthropic’s decision to donate Agent Skills to the Agentic AI Foundation](https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation). The announcement concerns a name, a logo, a GitHub organization, and a domain. The more interesting question concerns what travels with them: instructions meant to turn a capable general-purpose model into something that can do a particular job.

The format is modest. A skill is a folder containing a SKILL.md file, along with whatever scripts or reference material the task needs. Anthropic says it developed skills while working on Claude Code because Claude Code handled general knowledge work well but often lacked the procedural knowledge required by real business tasks.

It could know the subject without knowing the procedure. Anyone who has explained a familiar job to a new colleague will recognize the difference. The facts are only part of the work. There is also the order of operations, the exception that matters, the point at which proceeding would be a mistake.

Writing those things down makes them available to an agent. It also makes them available to other people, who can inspect, change, and carry them into their own work. Anthropic reports that more than 70 agent platforms have integrated the standard since its publication in December 2025. That is evidence of adoption as reported by its creator. It does not tell us how well any particular skill performs.

A saved talk, [“Building Great Agent Skills: The Missing Manual,”](https://www.youtube.com/watch?v=UNzCG3lw6O0) approaches that second question. Its speaker describes people assembling skills without knowing how to distinguish the useful ones from the disappointing ones. He argues that a shared rubric is missing, then proposes a checklist beginning with how a skill is triggered and how its instructions are structured.

That is one speaker’s diagnosis, not a census of every evaluation method. But it offers a useful question to carry back to the announcement: once instructions can travel, how do we judge what they teach?

Imagine a skill for preparing a weekly report. The file might explain where to find the numbers and how to arrange them. A better version might also explain what to do when a source is missing, which comparisons are misleading, and when to stop rather than fill a gap with plausible prose. Both could arrive in the same kind of folder. The difference would emerge in the work.

The donation gives the standard a new home. Anthropic says engineers from several companies will maintain it as individuals, and that people can continue to contribute without joining the foundation. Its proposed roadmap includes versioned releases, a formal proposal process, and security guidance.

Those proposals are worth following. They leave room for a separate discipline: testing the instructions against the tasks they claim to teach.

The folder can move from tool to tool. Whether it deserves to come along is still a question someone has to answer.