SfSkills: Salesforce skills for AI coding assistants

Grounded Salesforce skill packages, run-time agents, a requirement-to-build loop and an MCP server for Claude Code, Cursor, Codex and any MCP client.

SfSkills makes an AI coding assistant work on Salesforce the way a senior practitioner does. It is a library of 1,040 skill packages covering Apex, LWC, Flow, Agentforce, OmniStudio, integration, data, security, DevOps and declarative admin work; 70 run-time agents you call as slash commands in Claude Code; a beta loop that turns one business requirement into a verified, deploy-ready build; and an MCP server that gives any MCP client the same library plus read-only probes of your org. Nothing in this repository deploys to an org.

Contents

• What this is
• Sixty-second start
• What a session looks like
• What’s in it
• Why you can trust the output
• Limits
• Docs
• Contributing
• License

What this is

1. A skill library. Each package under skills/// is a SKILL.md (when to use it, a recommended workflow, and in packages written or revised since 2026-09-04 the questions to ask before configuring) plus four reference files and a checker script:

• references/examples.md: worked examples, with metadata XML where the topic has it
• references/gotchas.md: the platform’s non-obvious failure modes
• references/llm-anti-patterns.md: the wrong code an LLM reliably writes in this area, why it writes it, and the correct pattern (example)
• references/well-architected.md: the Well-Architected mapping and the official Salesforce sources the package rests on
• scripts/: a standard-library Python checker for the package’s known failure modes

An assistant that has read the package refuses the specific wrong pattern and can say which document