Combine Skills and MCP to Close the Context Gap — Pedro Rodrigues, Supabase

Key Concepts

• AI Agent Skills: Folders containing instructions, scripts, and resources that allow AI agents to discover and execute specific tasks.
• MCP (Model Context Protocol): An open standard that enables AI agents to connect to data sources and tools.
• Front Matter: Metadata (name, description) within a skill file that helps the agent decide when to load the skill.
• Skill.md: The primary instruction file for an agent skill.
• RLS (Row Level Security): A PostgreSQL feature that restricts data access; critical for security when agents interact with databases.
• Evals (Evaluations): Automated tests used to measure an LLM’s reasoning, tool-calling accuracy, and behavior.
• DDL (Data Definition Language): SQL commands used to define or modify database structures.

---

1. The Role of Agent Skills vs. MCP

The speaker, Pedro, an AI tooling engineer at Supabase, clarifies that while MCP provides the "tools" for an agent to interact with a system, Skills provide the "guidance" on how to use those tools correctly.

• The Problem: Agents often rely on stale training data, exhibit "laziness" (avoiding external lookups), and may ignore security best practices (e.g., bypassing RLS in PostgreSQL).
• The Solution: By providing a structured "Skill," developers can force the agent to follow optimized, secure workflows that are specific to their product.

2. Principles for Building Product Skills

Pedro outlines three core principles derived from his experience developing the Supabase agent skill:

1. Avoid Duplication: Do not duplicate existing documentation. Instead, point the agent to the "single source of truth" (the official documentation). Be persistent in instructing the agent to search these sources.
2. Prioritize Critical Information: If information is vital (e.g., security checklists), place it directly in the skill.md file. Agents are "lazy" and often fail to load or reference secondary "bundled" files.
3. Be Opinionated: Developers should encode their preferred workflows into the skill. For Supabase, this means a specific schema-management workflow:
   • Perform direct DDL operations on a dev/staging database.
   • Use the Supabase advisor to check for security/performance issues.
   • Generate the migration file only after the schema is validated.

3. Testing Methodology: Evals

To ensure the skill was effective, the team used Braintrust to run evaluations.

• Setup: They tested six scenarios across four different models (Claude 3.5 Sonnet/Opus, GPT-4o/mini).
• Conditions: Baseline (no tools), MCP only, and MCP + Skills.
• Findings: The combination of MCP + Skills consistently outperformed other conditions in "test completeness scores," proving that guidance is as important as tool access.

4. Real-World Application: Supabase Documentation via SSH

Supabase is experimenting with exposing documentation via SSH. The rationale is that agents are highly proficient at navigating file systems using Linux-based tools. By treating documentation as a file system, agents can navigate and retrieve information more reliably than through standard web-search interfaces.

5. Notable Quotes

• "The bottom line is not the context, it's the guidance."
• "If something can get skipped, it will be skipped." (Referring to the agent's tendency to ignore reference files).
• "Don't be afraid of guiding the agents on what workflows you think are the most effective."

6. Distribution Challenges

During the Q&A, Pedro acknowledged that there is currently no universal standard for distributing skills.

• Current State: Distribution is fragmented; some companies use specific packages (e.g., Vercel’s skills package), while others bundle skills directly into repositories (e.g., .claude or .cursor folders).
• Internal Strategy: Supabase currently packages skills within the relevant repositories, making them discoverable to anyone with access to the code.

Synthesis/Conclusion

The primary takeaway is that as AI agents become more capable, the bottleneck shifts from "can the agent do this?" to "does the agent know the best way to do this?" Developers should treat agent skills as a form of "executable documentation." By being opinionated, keeping critical instructions in the main skill.md file, and rigorously testing via evals, teams can significantly improve the reliability and security of AI-driven workflows.