Cortex Code Skills System

Reusable AI workflows for Snowflake development
March 9, 2026 Edition
1 What is a Skill?
A skill is a folder that teaches Cortex Code how to perform a specific workflow.
my-skill/ SKILL.md # Required: instructions + frontmatter scripts/ # Optional: Python, Bash executables references/ # Optional: docs loaded on demand assets/ # Optional: templates, icons
  1. Project.cortex/skills/
  2. Global~/.snowflake/cortex/skills/
  3. Remote — Git repos (auto-cached)
  4. Bundled — Ships with Cortex Code
$skill-name do something specific

Type $ in the prompt for autocomplete. Manage skills with /skill interactive UI.

--- name: snowflake-migration description: "Migrate tables between schemas. Use when: schema migration. Triggers: migrate, move tables." tools: ["snowflake_sql_execute", "bash"] ---

The description field drives auto-activation. Include trigger keywords.

2 How Skills Work
Skills use progressive disclosure to minimize token usage while maintaining deep expertise.
L1
YAML Frontmatter Always in context — triggers auto-activation
L2
SKILL.md Body Loaded when relevant — full workflow instructions
L3
Linked Files On demand — references/, scripts/, sub-skills
Composability Multiple skills load simultaneously. Design skills to work alongside others.
Portability Works across CoCo CLI and Claude Code via .claude/skills/ compatibility.
Conciseness Keep SKILL.md under 500 lines. Only add what the LLM doesn't already know.
MCP (Connectivity) Skills (Knowledge)
Connects to tools & services Teaches how to use them
What CoCo can do How CoCo should do it
Real-time data access Workflows & best practices
3 Building & Using Skills
Use the built-in $skill-development skill to create, capture, and audit skills.
$skill-development create — Scaffold a new skill from scratch
$skill-development summarize — Turn a session into a reusable skill
$skill-development audit — Check a skill against best practices
--- name: snowflake-migration description: "Migrate tables between schemas" tools: ["snowflake_sql_execute", "bash"] --- # Workflow 1. Discover source objects 2. Generate DDL 3. STOP — Get user approval 4. Execute migration 5. Validate row counts 6. STOP — Confirm results

STOP markers = mandatory user checkpoints before destructive actions.

analyzing-data streamlit semantic-view cortex-agent machine-learning data-governance cost-management dynamic-tables lineage authoring-dags dbt-verify trust-center