Bootstrap an Analytics Agent's Ontology from Your Existing Stack

Where this fits

You've been asked to "make our data agent-ready," but your warehouse is a sprawl of tables, dbt models, and dashboards nobody fully remembers. Manually documenting every column and metric would take weeks. The ontology-bootstrap skill turns what your stack already has—schema exports, dbt YAML, query logs—into a structured Cassis ontology tree, with evidence attached to every claim and unresolved questions filed for human review. You stay in control at four checkpoints, and the whole run works offline with no API keys.

Why agents benefit

• Evidence-backed context: Every table, metric, and join in the generated ontology carries a pointer to its source (schema comment, dbt test, dashboard definition), so an agent can trace why a claim exists.
• Explicit unknowns: Ambiguous terms and missing descriptions are collected as open questions instead of silent guesses—the agent knows what it doesn't know.
• Offline determinism: Scripts run locally without network or credentials, making the bootstrap reproducible in air-gapped environments and CI pipelines.
• dbt write-back without clobbering: Descriptions merge into existing schema.yml files, preserving any hand-written commentary—safe to run on a live project.
• Human-in-the-loop checkpoints: The process pauses at four decision points (domain boundaries, table groupings, metric definitions, join rules), so the output reflects your team's judgment, not just heuristics.

Practical scenarios

Onboarding a new analytics engineer

Instead of a week of shadowing, the engineer runs the bootstrap on a read-only warehouse snapshot. They get a domain-organized README tree with tables grouped by business area and open questions to ask the data team—cutting ramp-up time dramatically.

Preparing a warehouse for a natural-language BI agent

Your team wants to deploy an agent that answers "What's our churn by region?" The bootstrap produces metric YAML with definitions and source lineage, so the agent can reason about the right table and join paths without hallucinating.

Auditing a legacy dbt project before a refactor

The skill's output surfaces undocumented tables and ambiguous metric names as questions. Those questions become a backlog for the refactor, and the ontology serves as a living map during the migration.

Add it to your agent workflow

Install the skill into your agent's registry:

npx skills add GetCassis/ontology-bootstrap --skill ontology-bootstrap

Then invoke it with a pointer to your warehouse export and dbt project. Example prompt: "Bootstrap an ontology from schema.sql and the models/ directory, using the default checkpoints." The skill will generate a cassis/ tree with READMEs, YAML files, and a questions.md—all ready for your review.

Compare before adopting

Before committing, compare against alternatives like manual documentation, LLM-generated glossary tools, or other ontology frameworks. Check: does it attach evidence to every claim? Does it handle your dbt version and warehouse dialect? Is the output structured for agent consumption (not just human reading)? And does the four-checkpoint flow match how your team likes to review? The skill's offline nature and non-destructive write-back are strong differentiators, but verify they fit your CI and review cadence.

Why it is worth tracking

ontology-bootstrap is young (3 stars) but solves a painful, recurring problem: turning implicit data knowledge into explicit, agent-readable context. Its deterministic, offline-first approach and evidence-tracking design are quality signals that the authors care about correctness and auditability. Watch for community adoption and dbt-version compatibility updates. If you're building analytics agents, this is worth a trial run on a non-critical schema—the output will show you how much of your stack is already documented and what's missing. For more agent skills, see the OpenAgentSkill directory.