Skills

A skill is a Markdown document with instructions or knowledge that an agent loads when it needs it. Skills keep the system prompt short: instead of describing every table and every rule up front, the prompt lists the available skills with a one-line description each, and the agent loads the relevant ones with the load_skill tool.

Typical skills:

Schema skills — describe a business domain's tables, with SQL guidelines and example queries, so the agent writes correct SQL for the query tool.
How-to skills — instructions for a type of request ("how to explain a fee", "how to handle a dispute").
Shared query skills — general guidance, such as how to query the micro-database.

Skill File Format

A skill is a .html file with YAML frontmatter:

---
description: Deposit account (checking/savings) schemas - accounts, transactions, overdrafts, fees, ACH transfers - plus SQL guidelines and example queries
tables: DDA_Accounts, DDA_Transactions, DDA_Overdrafts, DDA_Fees, DDA_ACH_Transfers
---
The `### Schema` section below is generated at load time from the live data model.

## SQL Guidelines
- The schema represents a single customer — no need to filter by customer_id.
- Only reference fields that exist in the schema below.

## Example Queries
- "What is my total savings balance?" →
  `SELECT COALESCE(SUM(balance), 0) FROM DDA_Accounts WHERE account_type = 'Savings';`

Frontmatter key	Description
`description`	Required. Shown in the agent's skills catalog — the model decides whether to load the skill from this line alone, so state precisely what it covers.
`tables`	Optional. A comma-separated list of LU tables. When the skill is loaded, the live schema of these tables — as DDL, with table and column documentation from the LU's documentation MTables (`*_doc`) — is appended to the body. The skill therefore always matches the deployed schema.
`title`	Optional, informational.

If schema generation fails, or a table is not part of the LU schema, the static body is used as is — a misconfigured skill never breaks the agent. Document reference-only tables in the body text.

Where Skills Live

Location	Scope	How an agent gets it
`<LU>/AI/Agents/<agent>/Skills/`	Agent-private	Discovered automatically.
`<LU>/AI/Skills/`	LU-wide	Must be listed under the agent's `skills:`.
`SharedObjects/AI/Skills/`	Shared across Data Products	Must be listed under the agent's `skills:`.

LU-wide skills are the natural home for schema skills: one skill per business domain, reused by the main agent and by the sub-agent that specializes in that domain.

skills:
  - dda-schema                         # LU-wide skill, loaded on demand
  - credit-card-schema:
      auto_load: true                  # injected into the prompt from the start
  - microdb-query                      # shared skill

Loading: On Demand or Auto-Load

On demand (default) — the skill appears in the ## Available Skills catalog of the system prompt. The agent calls load_skill when the request needs it. The catalog is also the allow-list: the agent cannot load a skill that is not in it.
Auto-load — auto_load: true injects the skill body under ## Loaded Skills in the system prompt. Use it for a specialist sub-agent whose single domain schema is always needed.

When the query tool is enabled and the agent has skills, the prompt tells the agent to load every skill whose description covers the data it is about to query, and to answer a cross-domain question with a single JOIN rather than one query per domain.

Skills and Sub-Agents

A sub-agent sees only the skills it declares itself. Its skill names are resolved in its parent agent's Skills/ folder first, then among the LU-wide skills. Declare in each sub-agent exactly the skills of its domain.

Writing Good Skills

Make the description specific. "Credit card schema — cards, statements, payments, transactions, MCC categories" routes better than "Credit card data".
One domain per schema skill. Smaller skills load faster and cost fewer tokens.
Include example questions with their SQL. Examples teach the model your naming, joins and date handling better than rules.
State the scope of the data. For entity micro-databases, say that the data belongs to one customer, so the model does not filter by customer ID.
Use tables: instead of pasting DDL. The schema stays correct when the LU changes.
Remember to deploy. Skills are read from the deployed Data Product.

Next article: References and Planning