Skills are structured documents that give an Atlas AI agent domain-specific knowledge for a particular task or dataset. When a user’s question is relevant to a skill, the agent loads that skill and uses its guidance to answer. The agent does not load skills that are irrelevant to the current question, which keeps the agent’s context focused on what matters.
What skills contain
A skill tells the agent how to approach a specific type of question. A well-written skill includes these elements.
-
Trigger description. A short statement that tells the agent when to load this skill, helping it decide whether the skill is relevant to the user’s question.
-
Data model guidance. Instructions on which views to query and which properties are important. The agent can discover the schema at runtime, but explicit guidance reduces errors on frequently-used query patterns.
-
Vocabulary mappings. A mapping between the terms users are likely to use and the field values stored in CDF. Vocabulary mappings are one of the most effective ways to improve agent accuracy because ambiguous terminology is a common source of incorrect queries.
When the user says "high priority," filter on priority = 1.
When the user says "overdue," filter where plannedEndDate is in the past and status is not "COMPLETED".
When the user says "rotating equipment," include assets with tag prefixes P- (pumps), C- (compressors), and T- (turbines).
-
Tool sequencing instructions. The order in which the agent should call tools when answering questions in this skill’s domain. For example, when a question asks about sensor trends, the agent should query the knowledge graph to identify the relevant time series, retrieve data points, and then analyze them if the user asks about trends or anomalies.
-
Response templates. Instructions on what a complete answer looks like, including which fields to include and how to format the output.
-
Examples. A small set of diverse examples covering distinct question patterns. A few well-chosen examples are more effective than a long list of similar ones.
How skills relate to agent instructions
Agent instructions (the system prompt) define how the agent behaves in every interaction: its tone, language, safety boundaries, and output format defaults. Skills provide domain-specific knowledge that the agent loads on demand.
Keeping these separate means the agent’s context stays focused on what matters for the current question. When domain knowledge loads on every interaction regardless of relevance, the agent processes far more information than it needs and may miss important instructions. Domain knowledge only enters the context when it is relevant to the current question. Workflow-specific guidance, vocabulary mappings, and tool sequencing belong in skills, not in the agent’s general instructions.
Skills as reusable building blocks
A skill is not tied to a specific agent. A skill that teaches an agent how to navigate your asset hierarchy, interpret naming conventions, or query your maintenance data model can be attached to any agent that works with that data.
This makes skills a building block for your agent library. A data scouting skill that retrieves sensor readings, alarm history, and recent work orders for a piece of equipment is just as useful in a root cause analysis agent as it is in an equipment health agent. Writing skills as self-contained units, without references to a specific agent’s identity or purpose, makes them straightforward to reuse.
Keeping skills focused
A skill that tries to cover too many topics at once loses effectiveness. Each skill should focus on one domain or workflow step. A root cause analysis agent, for example, might use three separate skills: one for gathering relevant data, one for identifying correlations and anomalies in that data, and one for building a causal chain from symptom to root cause.
If a skill grows long enough to become difficult to read, consider splitting it into two focused skills with clear, distinct triggers.
Last modified on June 18, 2026
