The leveled-up data analyst stack:

CODING AGENT in your terminal (Claude Code, or Cursor's integrated agent) for new analysis work. You describe the analysis, the agent writes the SQL or Python, you review.

MAGE for any pipeline that runs more than once. v1's recommendation holds. Now you don't have to write the Python; the agent does, and you wire it into the DAG.

DEVII for any project that needs an API. Don't build a FastAPI by hand. Generate the API from your database, build your dashboard against it, save the time for the analysis work.

A FRONTEND TEMPLATE (Svelte or React) that you reuse across projects. We'll provide one in section 8.5. For one-off dashboards, Streamlit and Plotly Dash are still fine; for recurring/customer- facing dashboards, the proper frontend pays off.

A LOCAL MODEL for routine operations: SQL formatting, schema documentation, column-name conversions, simple data cleaning. The local model handles 80% of the volume; you save the frontier subscription for the hard analyses.

The new pipeline:

1. Receive an ad-hoc analytics request.
2. Talk through the request with the agent in plain English. Iterate until you have a spec.
3. Write the spec into the project's CLAUDE.md.
4. Generate the SQL/Python with the agent.
5. Verify (read the diff, run the query, check the numbers).
6. If it's recurring, wire it into Mage with a schedule.
7. If it needs a UI, scaffold the dashboard from your template.

What used to take a week now takes a day. What used to take a day takes an hour. The skill that took years to develop in v1, Python and SQL fluency, now compounds because the agent handles the typing and you handle the thinking.

================================================================================ PART 8 — TEMPLATE LIBRARY

This section provides starter templates for the three categories: spec- writing, project types, and agents. Copy them. Modify them. Use them.

When something in your project doesn't have a template, write one and contribute it back. (See section 9.2.)