Claude can now load a folder of standing instructions on demand, which turns a team's best prompts from copy-paste text into reusable equipment.
The gap between a rep who gets good output and one who does not is almost never the model — it is whether the instructions are any good and whether they get reused. Skills make a good instruction set something you install once and call by name, rather than something each rep re-pastes and quietly degrades.
A Skill is a folder. Inside it is a file called SKILL.md — a plain Markdown text file, the same format used for README files — plus any scripts or reference documents the task needs. Claude reads the folder when it decides the task is relevant, and ignores it the rest of the time.
That last part is the design. Rather than stuffing every instruction into one enormous prompt, Claude loads only the pieces a given job needs. Anthropic calls this contextual loading. It exists because context windows are finite and because instructions that are always present start competing with each other.
The short version: a Skill is a named folder of standing instructions that Claude picks up when it is relevant, instead of you pasting the same brief every time.
Anthropic released Agent Skills on 16 October 2025. A Skill packages instructions, scripts and resources into a folder that Claude loads when it judges them relevant, using the SKILL.md format.
Availability: Claude apps on Pro, Max, Team and Enterprise; the Claude Developer Platform via the Messages API and a /v1/skills endpoint; and Claude Code via plugins.
Format: SKILL.md, plain Markdown, bundled in a folder with supporting files.
Requirements: The Code Execution Tool beta environment. Team and Enterprise deployments need organisation-wide admin enablement.
Price: Not stated in the announcement — Skills are a capability of existing plans, not a separately priced product.
Regions: Not stated.
What it doesn't do: a Skill does not make Claude better at your job — it makes Claude consistent at whatever you wrote down. A vague Skill produces vague output faster. Anthropic is also explicit that Skills execute code and that you should install only from sources you trust, which is a real constraint on sharing them casually between teams.
SDR — Least affected this quarter. Skills reward a repeatable, documented motion, and most SDR work is already templated inside a sequencer you do not control. Worth watching, not worth stopping for.
AE — This is the one that changes your week. A Skill holding your MEDDIC definitions, your qualification bar and your objection library means Claude assesses every deal the same way — so the output stops varying with how carefully you wrote that day's prompt.
Manager — Skills are the first practical answer to coaching consistency. A shared Skill is an enforceable standard, which a shared prompt document has never been, because nobody reads the document twice.
RevOps — This is a governance surface and should be treated as one. Skills execute code, they need admin enablement on Team and Enterprise, and an unmanaged skill library will fragment exactly the way unmanaged report folders did.
Before you write a Skill, find out whether your qualification standard actually survives contact with a live deal. Run this against your last three opportunities and see whether the account briefs come back consistent.
Two minutes, once a week. What changed in AI, and what to run because of it.