What do the five most-starred Claude Code workflow repositories have in common? More than you would expect. A new community comparison of the top approaches - ranging from 32,000 to 98,000 GitHub stars - reveals that despite different branding and terminology, they all converge on the same core architecture: Research, Plan, Execute, Review, Ship.
The five workflows compared are Superpowers (98k stars), Everything Claude Code (87k stars), Spec Kit (79k stars), Get Shit Done (35k stars), and OpenSpec (32k stars). Here is what the comparison tells us about how serious Claude Code users actually work.
The CLAUDE.md File Is Your Most Important Config
Every top workflow emphasizes the CLAUDE.md file - the project-level instruction file that shapes how Claude Code behaves in your codebase. The consensus best practices:
- Keep it under 200 lines. One well-regarded setup (humanlayer) uses roughly 60 lines.
- In monorepos, use multiple CLAUDE.md files at different directory levels. Claude loads them from ancestor to descendant, so you can set project-wide rules at the root and team-specific rules in subdirectories.
- Wrap domain-specific rules in conditional tags so they do not get overlooked in long files.
- Keep your codebase clean. Partial framework migrations (half React, half Vue, for example) confuse the model's pattern selection.
Agent Teams and Parallel Development
The most advanced workflows use sub-agents and git worktrees (isolated copies of your repo on separate branches) to run multiple Claude Code instances in parallel. The key insight: separate contexts produce better results than cramming everything into one long conversation. Several workflows recommend using tmux (a terminal multiplexer that lets you run multiple terminal sessions) to manage parallel agent sessions.
Slash commands - reusable prompt templates stored in .claude/commands/ - show up in every workflow as the way to standardize repetitive tasks. The recommendation is to write the command description as a trigger condition ("when should this fire?") rather than a summary of what it does.
Prompting Philosophy
The workflows share a surprisingly hands-off approach:
- Paste the bug, say "fix," and let Claude work. Micromanaging the approach produces worse results.
- After a mediocre fix, try: "implement the elegant solution knowing what you know now" - this leverages the context Claude built during its first attempt.
- Prototype aggressively. Multiple workflows recommend generating 20-30 versions rather than perfecting specifications upfront.
- Use separate Claude instances to review plans, acting as a "staff engineer" code reviewer.
The newest Claude Code features getting traction across these workflows include scheduled tasks (recurring prompts that can run for up to 3 days), voice dictation with 20-language support, and the agent teams beta for coordinated parallel work.
The overall pattern - Command, Agent, Skill - provides a clean mental model: commands inject knowledge into your current context, agents operate autonomously in fresh isolated contexts, and skills provide progressive capability within shared contexts. If you are using Claude Code without at least a basic CLAUDE.md file and a few slash commands, you are leaving significant capability on the table.