What is a Skill?

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: What is a Skill? · Channel: LangChain · 01:32 · Watch on YouTube

Topics: AI agents, Prompt and context engineering, AI developer tools

Summary

Skills help prevent long-running agents from making mistakes caused by overloading their context. A skill is a folder of task-specific instructions, scripts, and reference documents; at startup, the agent sees only the skill’s name and description. When a task calls for it, the agent can load the relevant skill and its supporting resources, while leaving unrelated context out. This layered approach keeps the context window small until information is useful.

Key points

Agents can make mistakes when their context is overloaded.
A skill is a folder containing context for a particular task.
Skills can include instructions, scripts, and reference documents.
At startup, only each skill’s name and description are loaded.
The short description helps the agent decide when to use a skill.
Task-specific context can be kept out of the system prompt until needed.
An expense skill might include spreadsheet instructions, a receipt-upload function, or a link to platform documentation.
With skills, information is loaded in layers rather than all at once.
Skills are one way an agent harness can manage its context window.

Chapters

Start	Chapter	Summary
00:00	What skills are and why they help	Skills address context overload by storing task-specific information in folders that agents can access when needed.
00:30	Keeping task context out of the system prompt	A personal productivity agent can use skills for preferences and procedures instead of loading all that information into its system prompt.
01:00	Loading relevant information in layers	The agent opens an expense skill only when needed, and its details do not carry over to unrelated tasks; skills are one context-management tool among many.

Quotes

Skills are one way to fix that. — 00:00

This folder can include instructions, scripts, and reference docs. — 00:00

Putting all of this context in the agent system prompt would be a waste of tokens. — 00:30

With skills, information is loaded in layers, not all at once. — 01:00

context stays small until it’s actually useful. — 01:32

Concepts and entities

skill, context window, agent harness, system prompt, LangChain, LangChain Academy

Full transcript (original language)

00:00 Ever wonder why a long-running agent starts making mistakes halfway through a task? It’s usually not the model, but everything crammed into its context and overloading it. Skills are one way to fix that. So, a skill is a folder filled with the context an agent needs to complete a certain task. This folder can include instructions, scripts, and reference docs. To protect an agent’s context window, the entire skill isn’t loaded in its startup. Just the name and description. This is just enough for the agent to know that the skill exists and when to use it. Kind of like knowing the

00:30 summaries on the back of every book in a library without actually opening the book and reading it until it’s relevant. For example, let’s say I had a personal productivity agent that I customize with my own preferences. Things like, this is how I like to sound when I write emails. This is exactly how to submit an expense. This is how to distill information when I prep for meetings. Putting all of this context in the agent system prompt would be a waste of tokens. I’m not always writing emails and it could also slow down my agent’s response times. Instead, I collect this context into skills and provide the

01:00 agent with only a short description of each, so it knows which one to grab and when. Now, if I ask it to submit an expense on my behalf, it can open the submit expense skill, which contains instructions on how I like to format my spreadsheets. In this skill, I could also include a function to help it upload receipts or a link to the docs page of our expense management platform. When I move on to something else, that expense detail doesn’t follow me to the next task. It was only ever in context because this task called for it. With skills, information is loaded in layers, not all at once. So, context

01:32 stays small until it’s actually useful. Now, skills are just one of the many things that an agent harness can use to manage its own context window. If you’d like to learn more, you should check out our free deep agents course on LangChain Academy.