Skills: Encoding Professional Expertise

In October, Anthropic shipped Agent Skills, an open standard for packaging instructions, scripts, and reference materials into structured files that AI agents can discover and execute. The procedural expertise that used to live in people's heads or lie buried in internal wikis can now be codified and made portable, editable, and executable by agents, which will change how companies think about staffing and how workers think about leverage.

What is a skill?

In the AI context, a skill is simply a folder containing a structured markdown document that tells an agent how to perform a specific task. For example, reviewing a document, drafting an article, checking code, or following brand guidelines. You can see Anthropic's own Brand Styling skill here for reference.

How skills work

At the root of every skill sits a SKILL.md file that includes a metadata header – giving the skill a name and description – followed by instructions for executing the task. At startup, the agent reads only the skill's metadata, giving it enough context to know what the skill is and when to use it.

The description then acts as the skill's trigger. When the agent encounters a task matching a skill's description, it reads the full SKILL.md body, loads any bundled scripts or reference files included with the skill, and leverages that context to complete the task. In this way, an agent can have dozens of skills installed without burning context on instructions it doesn't need. Skills are only called when needed.

Example

Here's what that looks like in practice, with a trimmed-down SKILL.md file for vendor contract review:

---
name: contract-review
description: >
  Reviews inbound vendor agreements where the company is the customer.
  Covers SaaS subscriptions, professional services, and data processing
  agreements. Produces a structured memo with risk assessment and
  recommended negotiation points. Use when reviewing a vendor contract
  or preparing for contract negotiation.
---

# Vendor Contract Review — Startup Context

## Reader assumption

Output is read by a founder or ops lead, not a lawyer.
Plain English throughout. No legal jargon without explanation.

## Step 1: Classify the contract

Identify contract type. Note governing law, jurisdiction,
effective date, expiration date, and whether companion documents exist.

## Step 2: Extract core commercial terms

Produce a summary table: vendor, contract type, term, auto-renewal,
fees, payment terms, fee escalation, governing law.

Flag immediately if:

- Auto-renewal notice window is under 60 days
- Payment terms are under Net 30
- Fees are subject to uncapped escalation

## Step 3: Review clauses in priority order

**Priority 1 — Flag and escalate:**
Uncapped liability, IP ownership of customer data, unilateral
modification rights, mandatory arbitration.

**Priority 2 — Note and recommend negotiation:**
Data portability on termination, termination for convenience, SLA and
remedy, price escalation caps, indemnification.

**Priority 3 — Record only:**
Notice requirements, assignment restrictions, confidentiality survival,
force majeure.

## Step 4: Produce the review memo

Executive summary (3-5 sentences): risk level, primary issue,
recommendation — sign as-is, negotiate, or escalate to counsel.

Then flagged items with section references and recommended asks.

Why skills matter

While the above file is an example, the steps listed are the same steps an experienced contract manager may apply in practice. Scope the review, extract the commercial terms, prioritize the clauses by risk, and produce a structured memo for review. That type of process-driven work can now be captured in portable, editable skill files that can be executed by an AI agent.

If professional expertise and procedural knowledge can be codified and performed by agents, companies can scale that knowledge without needing to scale headcount and without locking that knowledge inside software that requires an engineer to change.

It also puts pressure on the large number of workers whose value comes from executing those procedural workflows. Instead, the leverage will shift to workers who can analyze a complex process, understand why it works, and orchestrate agents to run it. The orchestration becomes key.