Boris Cherny, the creator of Claude Code, shared his workflow on Twitter, showing himself running 5-10 parallel Claude Code sessions simultaneously, claiming to produce 50-100 PRs per week. The Hacker News discussion (139 comments) reveals a community divided between enthusiasm and skepticism. Key themes include questions about scalability, cost concerns, rate limiting for normal users, and the practical challenges of reviewing AI-generated code at scale. Many commenters note that Boris's unlimited access as an Anthropic employee makes his workflow unreproducible for average users. The discussion also surfaced persistent bugs in Claude Code (UI flickering, concurrency issues) and debates about whether high PR counts translate to meaningful productivity.
| Pros | Cons |
|---|---|
| Claude Code works well "out of the box" with minimal customization needed | Running 10+ parallel agents is unrealistic for most users due to rate limits and costs |
| The "&" prefix allows "teleporting" tasks to cloud for background processing | Can be painfully slow on large codebases (20+ minutes for basic tasks) |
| Multiple instances work well as "state holders" rather than true parallelism | UI flickers badly in VSCode terminal, tmux, and some other terminals |
| Good for small, well-defined tasks (CSS fixes, component extraction, API endpoints) | Persistent concurrency bugs that developers admit they "can't fully eliminate" |
/resume command lets you switch between conversation states easily |
Every PR still requires human review - this becomes the bottleneck |
| Sub-agents and skills system provides conditional instruction execution | Claude Code can produce subpar code, especially on logic-dense tasks |
| Useful for brainstorming and read-only research sessions | Boris has unlimited tokens as an Anthropic employee - unrepresentative experience |
| Can run tests and CI/CD automatically before human review | High costs for heavy users (reports of $1k/day credit usage with poor results) |
| "Ultrathink" mode now enabled by default for better reasoning | 50-100 PRs/week claims are questionable - likely small/trivial changes |
| Helpful for CRUD apps and typical full-stack JS work | Tends to fix issues incorrectly by rolling back to older library versions |
| Can integrate with Playwright MCP for frontend browser viewing | Model quality varies; tools and pricing constantly change |
| Works well when you have many small, independent tasks | Hard to build sustainable institutional knowledge with constantly changing AI |
| Background task handoff between local and web Claude Code | Claude Desktop has bugs: loses chats, no easy context extension, slow |
| Plan mode helps create better specs before execution | Workflow only works for small features, not complex problems |
| Can combine multiple instances for different parts of codebase | Produces "so many errors" for some users on complex tasks |
| Some report it's more stable and faster than initial releases | iOS app constantly throws "unknown errors" |
- Use "&" prefix to send tasks to Claude Code web (background/cloud mode) - frees up local terminal
- Use
/resumeto switch back to previous conversation states without losing context - Keep a Google Doc open for notes while jumping between Claude tabs
- Use Git worktrees to run parallel sessions on the same repo without conflicts
- Run 2-4 instances max - beyond that, human working memory becomes the bottleneck
- Use instances as state holders, not for extreme parallelism
- Start in plan mode - go back and forth until you get the spec right
- Create a
plan.mdfile before execution for complex features - Use shallow clones when working on multiple branches in parallel
- Don't allow LLM to make implicit decisions - confirm architecture choices explicitly
- Planning phase can take 1+ hour but execution is much faster when done properly
- Write clear, specific instructions - Claude works better with detailed requirements
- Build feedback loops for self-validation (tests, linting, CI/CD)
- Have one Claude write code, another Claude review it (or use Codex for reviews)
- Run bake-offs between Claude and Codex - same prompt, separate environments, score results
- Always comb through PRs and test locally - don't trust blindly
- Use test-driven approach - instruct Claude to write and run tests to catch its own bugs
- Put "ultrathink" in CLAUDE.md to make extended thinking the default (though now it's default anyway)
- Use skills for conditional instructions that agents can invoke contextually
- Slash commands = skills (used interchangeably, with slash commands becoming callable in future releases)
- Skills only load name/description initially - full content loads when invoked (context efficiency)
- Use a native terminal (Ghostty, iTerm2) instead of VSCode terminal to avoid flickering issues
- Avoid tmux if experiencing flickering problems
- Try agentastic.dev for Git Worktree + Ghostty integration (macOS)
- Use clodhost.com for web interface to run multiple sessions (free)
- Use Beads for ticket-style task management with agents
- Ask Claude to do research first, then ask clarifying questions, then execute
- Use one terminal for small nitpicks - quick fixes you notice while reviewing bigger work
- Dedicate one terminal for running existing tests repeatedly
- Dedicate one terminal for codebase research/questions
- If context gets corrupted (connection drops), you may need to manually edit internal JSON state files
- Start new conversations for clean context when approaching context limits
- Don't reshove everything blindly when requirements change - plan first
- Check if Claude is rolling back library versions when encountering issues
"I don't need 10 parallel agents making 50-100 PRs a week, I need 1 agent that successfully solves the most important problem."
"Multiple instances of agents are an equivalent to tabs in other applications - primarily holders of state, rather than means for extreme parallelism."
"The program is becoming buggier and less reliable over time"
"Do not allow the LLM to make any implicit decisions, but instead confirm with the user"
"Having one LLM review the PRs of another is actually useful as a first line filter"
- Reddit thread with more context: Boris's follow-up answers
- Alternative workflow with sub-agents: solatis/claude-config
- Git worktree tool: agentastic.dev
- Web interface for parallel sessions: clodhost.com
- Claude Code Skills documentation: code.claude.com/docs/en/skills