Skills

A skill is prose an agent loads only when it decides it needs it: a SKILL.md file the model reads on demand, instead of a paragraph folded permanently into instructions.

from agentdeck import Agent

triage = Agent(
    name="Triage",
    instructions="Route support tickets. Load the refund-policy skill before approving a refund.",
    skills=["refund-policy"],
)

skills=[...] names skills by directory, resolved against the Deck's skill roots at build(). A name the catalog does not hold fails there, naming the agent and what is available.

SKILL.md

.agentdeck/
└── skills/
    └── refund-policy/
        └── SKILL.md

---
name: refund-policy
description: How to evaluate a refund request - use before approving or denying one.
---

Approve refunds under $50 without escalation. Anything over $50 needs a supervisor tag.
Never approve a refund for an order older than 90 days.

The directory name is the skill's name; name: in the frontmatter must match it if given, and description: is required. Both missing or mismatched fail Deck.build() naming the bundle. Only description reaches the agent's own instructions up front; the model calls load_skill("refund-policy") to read the body, which is why a skill is not a tool call the model can fake its way past.

Deck.from_project() scans ./.agentdeck/skills/ automatically. Building a Deck directly instead, point skills= at the same directory (or several, or a Skills(...) object):

from agentdeck import Deck

deck = Deck(agents=[triage], skills="./.agentdeck/skills")

See Agent with a Skill for a full running example, tools and all.