Analyzing Insights Across Teams

Analyzing insights across teams

system.* entity tables (e.g. system.insights) are scoped to the current project, and the generic execute-sql guidance says other teams' data is inaccessible. For the dogfood project (US project 2) that is not the whole story: production Postgres tables are replicated into the project's data warehouse, so cross-team entity metadata is queryable with posthog:execute-sql. Do not stop at system.insights when the question spans teams.

This skill is deliberately repo-local (.agents/skills/): it documents PostHog's internal dogfood setup, applies only to agents working in this repo, and must not move into the packaged products/*/skills/ bundle that ships to every team.

Synced tables

• Underscore aliases (e.g. postgres_posthog_dashboarditem) point at the same synced data.

• These are replicas of the Django tables in this repo (posthog_dashboarditem backs the Insight model), so rows span every team; team_id is the scoping column.

• More prod tables than these are synced. Before concluding cross-team data is inaccessible, check the catalog:

  SELECT table_name, description
  FROM system.information_schema.tables
  WHERE table_type = 'data_warehouse' AND table_name ILIKE '%postgres%'
  

Workflow

1. Confirm columns before projecting — synced schemas drift with the Django models:

   SELECT column_name, data_type
   FROM system.information_schema.columns
   WHERE table_name = 'postgres.posthog_dashboarditem'
   

2. Query with posthog:execute-sql, filtering or grouping by team_id. Example — most active teams by insights created in the last 30 days:

   SELECT team_id, count() AS insights_created
   FROM postgres.posthog_dashboarditem
   WHERE NOT deleted AND saved AND created_at >= now() - INTERVAL 30 DAY
   GROUP BY team_id
   ORDER BY insights_created DESC
   LIMIT 20
   

   Join postgres.posthog_team on id = team_id for team names only when the output stays on an internal surface (see below).

3. Remember the sync lag: these are periodic replicas, not live reads — fine for analysis, not for "right now" state.

Output handling (required)

Rows in these tables are customer data: team names, insight names, descriptions, and queries.

• Never put customer team names, insight titles, or other row-level metadata on public surfaces — PR titles/descriptions, commit messages, issues, code comments, or uploaded screenshots. Aggregates and team_id-level figures without names are the ceiling for public copy.
• Keep named results in the private conversation, internal docs, or auth-gated links.
• Access is gated by membership in the internal dogfood project. If a query fails with a permissions error, report it and stop — do not look for another route to cross-team data.

Related

• For cross-team event/analytics data (not entity metadata), see the querying-production-databases-via-metabase skill instead.