Dbt Agent Readiness

dbt-agent-readiness

Find the places your AI analyst will return a plausible — but wrong — answer.

`dbt-agent-readiness` is a Claude Code skill that audits a dbt project for ambiguous metrics, unsafe joins, misleading documentation, invalid columns, and unclear model grain. It is built for dbt teams piloting AI analysts, copilots, or internal data agents.

The audit produces one evidence-backed Markdown report containing:

• what an agent will get wrong today
• what still needs runtime verification
• which models are safe to query
• what to fix first

It does not query your warehouse, require warehouse credentials, or modify your dbt models and configuration.

Quick start

Clone the skill into your personal Claude Code skills directory:

git clone https://github.com/GetCassis/dbt-agent-readiness ~/.claude/skills/dbt-agent-readiness
python3 -m pip install -r ~/.claude/skills/dbt-agent-readiness/requirements.txt

Then ask Claude Code:

Run the dbt-agent-readiness skill on /path/to/dbt/project

The report is written to `{project_path}/dbt-agent-readiness.md`.

Requirements

• Claude Code with Skills support
• Python 3.9+
• pyyaml and `sqlglot`, installed by the command above

Running `dbt compile` in the target project before the audit is recommended. It lets the skill resolve generated columns from macros such as `dbt_utils.star`, `SELECT *`, and Jinja loops. When compiled SQL is unavailable, checks that cannot be supported confidently are suppressed instead of reported as findings.

See what it finds

On the bundled 10-model test project, the audit finds four concrete ways an agent can fail:

Failure	What the agent gets wrong
Inconsistent entity names	Misses rows when joining customer_id, cust_id, and user_id
Competing revenue marts	Picks between two plausible models with different grains
Misleading descriptions