Skip to content

cross-agent-workflows.md

Latest commit

 

History

History
176 lines (135 loc) · 4.83 KB

cross-agent-workflows.md

File metadata and controls

176 lines (135 loc) · 4.83 KB

Cross-Agent Workflows

How to use pro-workflow patterns across Claude Code, Cursor, Codex, and other AI coding agents.

The Layered Approach

The most productive teams in 2026 don't choose one tool — they layer them:

Layer Tool Strength
Editor Cursor / VS Code + Copilot Tab completions, inline edits, visual diffs
Terminal Claude Code / Codex CLI Deep reasoning, multi-file changes, CI/CD
Background Cursor Background Agents Long-running tasks, PR creation
Review Claude Code reviewer agent Security audit, code quality

Common Setup

Cursor for daily coding (tab completions, inline edits)
  └── Claude Code in Cursor's terminal (hard problems, refactors)
      └── Subagents for parallel exploration

The overlap in capabilities is worth the combined cost.

Configuration Mapping

Claude Code → Cursor

Claude Code Cursor Equivalent
CLAUDE.md .cursorrules or .cursor/rules/*.mdc
.claude/settings.json .cursor/settings.json
.claude/agents/*.md .cursor/agents/*.md (v2.4+)
.claude/skills/*/SKILL.md .cursor/skills/*/SKILL.md (v2.4+)
.claude/commands/*.md Built-in / commands
hooks.json Not supported (use rules instead)
.mcp.json .cursor/mcp.json

Claude Code → Codex CLI

Claude Code Codex Equivalent
CLAUDE.md AGENTS.md
Settings codex.json
Agents Not supported
Skills Not supported
Hooks Not supported
MCP Supported via config

Claude Code → Gemini CLI

Claude Code Gemini Equivalent
CLAUDE.md GEMINI.md
Settings gemini.json
Agents Not supported
Skills Not supported

Pro-Workflow on Each Agent

Claude Code (Full Support)

Everything works: commands, agents, skills, hooks, rules, contexts, MCP config.

/plugin marketplace add rohitg00/pro-workflow
/plugin install pro-workflow@pro-workflow

Cursor (Skills + Rules)

Uses skills (.mdc rules) instead of hooks. Same patterns, different enforcement.

/add-plugin pro-workflow

What loads:

  • 9 skills (pro-workflow, smart-commit, wrap-up, learn-rule, etc.)
  • 6 rules (quality-gates, atomic-commits, context-discipline, etc.)
  • 3 agents (planner, reviewer, scout)

Other Agents (via SkillKit)

npx skillkit install pro-workflow
npx skillkit translate pro-workflow --agent codex
npx skillkit translate pro-workflow --agent gemini-cli
npx skillkit translate pro-workflow --agent windsurf

SkillKit translates the SKILL.md format to each agent's native format.

Shared Patterns That Work Everywhere

These patterns are agent-agnostic:

1. Self-Correction Loop

Works in any agent that reads a memory file. Add to CLAUDE.md / .cursorrules / AGENTS.md:

When corrected, propose a rule. After approval, append to LEARNED section.

2. Plan Before Multi-File Changes

Universal pattern. Every agent benefits from planning mode:

  • Claude Code: Shift+Tab to plan mode
  • Cursor: Agent mode with planning prompt
  • Codex: Start with "plan this before implementing"

3. Quality Gates Before Commit

Add to any agent's rules:

Before committing: run lint, typecheck, and tests. Fix failures before proceeding.

4. Context Management

Every agent has finite context. The discipline is the same:

  • Read before edit
  • Compact/clear at task boundaries
  • Delegate heavy exploration to subagents
  • Keep <10 MCP servers, <80 tools

5. Session Handoffs

End sessions intentionally. Capture state for the next session:

When ending a session:
1. List what's done, in progress, and pending
2. Note any decisions made and why
3. Write a one-line resume prompt for next session

Background Agent Patterns

Claude Code

Ctrl+B          # Send to background
claude -w       # Worktree for parallel session

Cursor

Background agents run on isolated VMs. They can:

  • Work on separate branches
  • Open PRs for review
  • Record video of their work
  • Run indefinitely

When to Use Each

Task Claude Code Cursor
Quick exploration Subagent (in-process) Tab in sidebar
Long refactor Worktree session Background agent
PR creation gh pr create in session Background agent auto-PR
Code review Reviewer agent Inline review comments

MCP Server Sharing

MCP servers work across both Claude Code and Cursor. Share a single .mcp.json:

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp@latest"]
    },
    "playwright": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-playwright"]
    }
  }
}

Place at project root. Both tools discover it.