Last Updated: 2026-04-12 This document tracks which AI development capabilities exist on each platform (Cursor IDE, Claude Code, GitHub Copilot) and whether the gaps are intentional (platform-specific) or represent work to close. Canonical quick reference: AGENTS.md
Hierarchy:
constitution.md → AI_GUIDE.md → AGENTS.md → .cursor/rules/ → core-specific AGENTS.md
Platform Overview
Capability Matrix
Agents / Specialists
Cursor total: 24 agents | Claude total: 6 agents | Copilot: 0 (path-scoped instructions only)
Skills / Workflows
Cursor total: 23 skills | Claude total: 5 skills | Copilot: 0
Rules / Pattern Files
Cursor total: 36 rules | Claude total: 4 path-scoped | Copilot: 5 path-scoped
Commands / Workflows
Cursor total: 73 commands | Claude total: 1 command | Copilot: 0
Hooks & Automation
Policy: Shared vs Platform-Specific
Shared (All Platforms Must Have)
All platforms must convey these policies to AI (via rules, instructions, or agent prompts):- Tenant isolation:
organization_idon all mutations - No PHI in prompts, logs, or error messages
- No direct core-to-core imports — use
@/platform/* - RLS on all new tables
sanitizeErrorMessagefor user-facing errorsuseCurrentUserfor auth (not ad-hocsupabase.auth.getUser())staleTime: 5mandgcTime: 10mon alluseQuerycalls- No hardcoded state Medicaid rules — use PF-96 jurisdiction profiles
- No inline
pf_user_role_assignmentsin edge functions — useverifyOrgAccess()
.claude/rules/ path-scoped) | Copilot (⚠️ partial — AGENTS.md + some instructions)
Intentionally Platform-Specific
- Cursor only: Full agent fleet (24 agents), 73 commands, skills, MCP servers, IDE hooks
- Claude only: SessionStart reinject hook, file organizer skill, notification hook
- Copilot only: Shared-UI path-scoped instructions (
.github/instructions/shared-ui.instructions.md)