D4 · Skills (progressive disclosure)
Family: Context and memory
Folders of instructions of which only the name and description are loaded at the start; the rest comes on demand.
Principle
A skill is a folder of instructions, scripts and resources. At startup, only its name and description are in the context. If the agent judges the skill relevant, it reads the full SKILL.md file, then the supporting files only when needed.
The problem it solves
Putting every instruction and every piece of know-how in the system prompt bloats it, while each task uses only a small part.
When to use it
- An agent must master many procedures (documents, reviews, deployments) of which only one is used at a time.
- You want to share and version team know-how as plain folders of files.
When to avoid it
- The instructions are short and always needed: put them directly in the prompt.
- Loaded content piles up over a multi-domain request: prefer a context-isolation subagent (D5).
Common pitfalls
The anti-patterns that misuse this pattern, best established first. Each sheet describes the symptoms, the public incidents and the remedies.
Instructions
- AP21 · Agent instruction bloat — An instructions file that grows with every incident, until important rules get ignored.
Security and permissions
- AP67 · Unverified tool supply chain — MCP servers, extensions and skills installed with no vetting or pinning, and the agent’s full trust.
- AP63 · Security policy left to the prompt — The “don’t touch production” rule only lives in the prompt; nothing enforces it when a tool is called.
- AP65 · Agent that rewrites its own configuration — The agent can write to the files that set its permissions: an injection no longer bypasses approval, it switches it off.
Oversight, operations and governance
- AP79 · Shadow and orphaned agents — Agents that are published, shared or forgotten, and that nobody monitors, caps or can revoke.
Synonyms by vendor
- Agent Skills (Anthropic (oct. 2025 ; standard ouvert en déc. 2025))
- Skills (LangChain (multi-agent))
- Bibliothèque de compétences (ancêtre académique) (Voyager (Wang et al., TMLR 2024))
Neighbouring patterns
Sources
- Equipping agents for the real world with Agent Skills — Anthropic (B. Zhang, K. Lazuka, M. Murag), 2025-10-16
- Multi-agent — LangChain (documentation v1)
- Voyager: An Open-Ended Embodied Agent with Large Language Models — G. Wang et al. (TMLR), 2024-03
Content reviewed on 2026-10-02