Descriptions:
Pedro Rodrigues, AI tooling engineer at Supabase, uses his AI Engineer conference slot to officially announce the Supabase agent skill — a structured document package that gives AI agents the product-specific knowledge needed to work correctly with Supabase’s backend-as-a-service platform. The announcement is made live on stage, complete with a live tweet, marking the public release of a resource Rodrigues spent months developing.
The core problem the skill addresses: agents operating solely on training data make predictable Supabase-specific mistakes. The clearest example shown involves Row Level Security — when an agent creates a PostgreSQL view over an RLS-protected table without explicitly setting `security_invoker = true`, the view silently bypasses access controls and exposes data. Tested with Claude Sonnet 4.6, an agent equipped with the skill caught and fixed this; one relying only on MCP tool access did not.
Rodrigues distills the team’s experience into three principles for building product skills. First, avoid duplicating existing documentation — point the agent to canonical sources rather than restating them. Second, treat anything skippable as if it will be skipped; critical security rules and non-negotiable workflows must live in the root `skill.md` file, not buried in optional reference files, because agents are reliably lazy about loading secondary files. Third, be opinionated: you know your product’s optimal workflows better than any general-purpose model, and the skill is the right place to encode them. The Supabase example encodes a specific schema management workflow — run DDL freely in dev, check with the advisor for security and performance issues, then generate the migration file.
📺 Source: AI Engineer · Published May 15, 2026
🏷️ Format: Keynote Launch
