ChatGPT Work's data agent: build a data-context skill from Databricks, Slack and Drive

Step-by-Step Breakdown

1. Connect the tools you'd use yourself.

   The framing: give the agent whatever context you would give a new analyst. In ChatGPT Work, open the Plugins tab (top left), click + on each plugin and sign in. On Business and Enterprise plans an admin may have to enable a plugin first.

2. Point the data plugin at your sources.

   The demo uses three: data tables in Databricks, an analytics channel in Slack, and a Google Drive document with metric definitions.

3. Generate the data-context skill.

   The data plugin reads across those sources and drafts a context skill: what business terms mean (a site viewer is not a site creator), which sources and activity count, which reporting period to use, and how to compare consistently.

4. Have the data team review it, then package it as a plugin.

   Someone who knows the data should edit and approve the skill. Only then is it packaged into a plugin and shared, so everyone's questions run on the same definitions.

5. Use it: question → evidence → plan → dashboard → alert.

   The overview clip shows the full loop. Ask from your phone, then open the evidence (metric definitions, underlying data, the SQL behind each number), check Slack and Drive for the reasons, assign owners and share the plan in Slack, turn the analysis into a shared dashboard, and set an alert such as "weekly creator adoption drops versus last week". You define how big a drop matters, when to check, and who hears about it, and you review all of that before scheduling it.

Gotchas & Caveats

• The quality of every answer depends on the reviewed context skill. An unreviewed auto-generated one will confidently repeat whatever ambiguity your Slack channel contains.
• Always open the SQL behind a number before acting on it. The product puts it one click away for this reason.
• These are vendor demos with a prepared dataset. Expect more work on a real warehouse.

Key Takeaways

• For a data agent, the context layer does the same job CLAUDE.md does for a coding agent.
• Build it from the sources your analysts already trust, and have a human review it before sharing.
• Package reviewed context as a plugin so it is shared and consistent rather than personal.