Agent Skills
1. Agent Skills
Hi, welcome back. How much time do you spend every week copy-pasting the same multi-step prompts for routine data tasks? Re-explaining context? Resetting guardrails? And hoping the output is as good as it was yesterday? What if your team's best practices, multi-step workflows, and executable scripts were pre-packaged and ready to run on demand? With guaranteed consistency, good performance, and zero time wasted on repeat prompts? Agent skills help with that. In this module, you will learn about agent skills, what they are, how they are structured, and how to invoke them. Then, we will go over use cases and give an introduction to some of the built-in CoCo skills that are useful for Snowflake data engineering tasks. Next, we will review how to create a custom skill and how to write them in an effective way, following best practices. Finally, we will talk about Snowflake's skills repository and how it works so you can publish skills to be shared with others. Skills extend CoCo with reusable, pre-packaged knowledge and workflows for specific tasks. A skill is a self-contained folder with a SKILL.md file describing what the skill does and the steps the agent should follow when it is invoked. You can use the skills packaged with CoCo Desktop, install community skills from a Git repository, share skills with your team, or author your own. All skills are managed from the Agent Settings panel. To open it, click the Agent Settings tab in the left sidebar. In the Agent Settings panel, select the Skills category from the left sidebar. This is the recommended place to browse, add, update, favorite, disable, and remove skills. Once a skill is registered, the agent can invoke it in two ways, automatically by matching your prompt against each skill's description. If your message clearly matches a skill, for example, "create an Iceberg table" matches the iceberg skill, the agent runs that skill's instructions before doing anything else. Alternately, a skill can be invoked explicitly by mentioning the skill in chat. Type / in the chat input and pick a skill from the picker. You can type some characters to find the skill or reference it by name, for example. Use the Introduce CoCo Desktop skill. When the agent invokes a skill, you see a Using Skill Name indicator in the chat transcript. Click the skill name to open its SKILL.md and inspect what the agent loaded. You can chain several /skill commands in a single message, and CoCo runs each one. There are many data engineering use cases, for example, migration, analysis and data profiling, building pipelines, debugging and monitoring, to name a few. You can use agent skills to help ease the many data tasks and ensure you have successful outcomes consistently. CoCo knows your Snowflake environment, context and company rules and can really help. CoCo will become your new favorite coworker. CoCo also includes a large set of built-in skills which enhance the model's general knowledge of Snowflake features. Here is a brief summary of those built-in skills for data engineers. For example, you can use a skill to deploy a native dbt project on Snowflake, or use the Lineage skill to analyze data lineage and dependencies. Agent skills are used for implementing repeatable, multi-step workflows and are made up of both instructions as well as scripts and tools. Skills are supported by most AI coding agents, and this emerging open standard is documented at the website agentskills.io. One key thing to keep in mind when getting started with skills is that the model is already intelligent and you only need to tell it what it can't infer. So how do you build a custom agent skill, and what do they look like? Like most extensibility features in generative AI coding tools, you define them primarily in natural language, via markdown files. And the core set of instructions for a skill is contained in the SKILL.md file. Here's a diagram that shows the folder structure of a skill. At the minimum, a skill consists of a folder with a single file named SKILL.md, which follows a few conventions. But the other really important part of a skill is the ability to package reusable code scripts in the /scripts folder, which help make a skill very consistent and reproducible. Here's an example. The SKILL.md file consists of an initial block of YAML frontmatter, followed by your own markdown instructions that the agent reads when the skill is invoked. Only the name and description fields are required. The model uses the description to decide when to invoke your skill. Writing well-designed skill bodies is the single best way to make your AI coding agents deterministic, fast, and secure. By defining clear skills up front, you eliminate hallucinations, protect your context window, and guarantee production-ready results. Every high-impact skill rests on three architectural pillars. Workflows that enforce a strict discover, plan, execute, and verify loop. Stopping points that act as circuit breakers after failed retries or trigger human approval. And guardrails that lock down forbidden paths and preserve critical logic. To execute these pillars effectively, keep two essential best practices in mind. First, enforce a strict 500-line limit on your core skill file to keep context lean. Offload deep documentation to reference files and heavy tasks to local executable scripts. And second, use precise third-person trigger keywords in your frontmatter so the agent activates the skill only when intended. For example, use present tense third-person verbs like generates, executes, parses, refactors, or validates. As you can see in this example, where the verb generates is used. When you treat your skill bodies as deterministic blueprints, your AI agents become reliable coding partners. In other words, focus on precision, structure, and repeatability. Then, skills turn unpredictable model behavior into solid engineering outcomes. The Skills Catalog is Snowflake's built-in registry for sharing and discovering skills across your organization. You can browse for skills in the Catalog using the Browse Skills Catalog link in the Skills view. To import a skill from the Catalog, in the Shared section, select the skill, then click the plus button and choose Add from Skills Catalog. And you can publish a skill to the Catalog. However, skills alone have limitations, so we will talk more about plugins and how they address those limitations in the next module. Now that you have a better understanding of agent skills, make sure you do the next hands-on exercise lab where you will create your own custom skill. You will then use the new skill in the next module's exercise.2. Let's practice!
Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.