claude-cracks-the-whip

Claude Code is the boss. Other AI agents are the workforce. Whip goes crack.

A Claude Code skill that turns Claude into a no-nonsense foreman — handing out assignments to AI coding agents (Codex, Gemini, Kimi, whoever shows up), watching them work, inspecting their deliverables, and sending sloppy work back.

The Problem

You're using Claude Code. It's great at planning, reviewing, and understanding your codebase. But when you have 5 independent tasks, doing them one-by-one is slow.

Meanwhile, other AI agents are great at heads-down grunt work but terrible at planning.

What if Claude Code could just... boss them around?

Claude Code (the boss)
    |
    |-- "You. Build this component. Now." -----> Agent 1 (at their desk)
    |-- "You. Write these tests. No excuses." -> Agent 2 (at their desk)
    |-- "You. Add this endpoint. Stay in lane." -> Agent 3 (in the back room)
    |
    v
  Inspects all deliverables. Sends sloppy work back. Ships the rest.

What It Sounds Like

> crack the whip on these 3 features

Rolling call... Codex is on the roster. tmux is available. Good.

Cracking the whip — deploying 3 agents.

Agent 1 is at their desk — working on UserAvatar component.
Agent 2 is at their desk — working on auth middleware.
Agent 3 is at their desk — working on API endpoints.

...

All agents have clocked out. Inspecting deliverables.

Agent 1: Clean work. Passes inspection.
Agent 2: Cut corners on error handling. Sending them back with corrections.
Agent 3: Clean work. Passes inspection.

Correction sent to Agent 2. Waiting for resubmission...

Agent 2 resubmitted. Passes inspection this time.

All work signed off. Ready for your review.

Features

• Roll call — auto-detects which agents are available (CLI tools, MCP servers)
• 3 work modes: at their desk (tmux), in the back room (background), phone it in (MCP)
• Instant notifications via tmux wait-for — agents tap the boss on the shoulder when done (polling fallback for non-tmux)
• Work logs via tee — Claude reads the agent's full output for smarter reviews, not just git diff
• Inspection loop — reads work logs, reviews diffs, runs quality checks, sends sloppy work back
• Agent-agnostic — any AI that takes orders can join the crew

Install

Option 1: Copy

cp -r skill/ ~/.claude/skills/claude-cracks-the-whip/

Option 2: Symlink (for development)

ln -s "$(pwd)/skill" ~/.claude/skills/claude-cracks-the-whip

Option 3: One-liner

git clone https://github.com/YOUR_USERNAME/claude-cracks-the-whip.git /tmp/cctw && \
cp -r /tmp/cctw/skill ~/.claude/skills/claude-cracks-the-whip && \
rm -rf /tmp/cctw

After install, restart Claude Code. Triggers on: "use codex", "dispatch to codex", "crack the whip", "dispatch agents".

How It Works

1. Roll Call

Claude checks the roster:

Codex? Present.
tmux? Present.
In a tmux session? Affirmative.

Full crew available. Deploying via tmux panes.

2. Write Assignments

Claude writes no-nonsense specs. Every assignment starts with:

You have one job. Do it now. No questions. No brainstorming.
No proposals. Just make these exact changes.

Then gives clear deliverables, quality checks, and boundaries. Stay in your lane.

3. Deploy

Agents get sent to their desks (tmux panes) with full logging and completion signals:

# Agent works in a visible pane, output logged, signals when done
tmux split-window -h "cat /tmp/assignment-1.txt | codex exec --full-auto -C '$REPO' - 2>&1 | tee /tmp/worklog-1.txt ; echo \${PIPESTATUS[0]} > /tmp/report-1 ; tmux wait-for -S agent-1-done"
tmux split-window -v "cat /tmp/assignment-2.txt | codex exec --full-auto -C '$REPO' - 2>&1 | tee /tmp/worklog-2.txt ; echo \${PIPESTATUS[0]} > /tmp/report-2 ; tmux wait-for -S agent-2-done"

4. Wait & Inspect

Claude blocks on tmux wait-for — no polling loops. Agents tap the boss when done:

# Blocks until agent signals (instant, no polling)
tmux wait-for agent-1-done
tmux wait-for agent-2-done

Then inspects:

• Reads work logs (what did they actually do, any errors, their reasoning?)
• Reviews git diff (what changed?)
• Runs quality checks (type-check, lint, tests)
• Sends sloppy work back with correction assignments
• Signs off on clean work

Work Modes

| Mode | Metaphor | When | Requires | |------|----------|------|----------| | tmux panes | At their desk | You want to watch them work | tmux + in session | | Background | In the back room | No tmux, or headless | Just the agent CLI | | MCP | Phone it in | Need back-and-forth | MCP server configured |

Agent Roster

| Agent | CLI | MCP | |-------|-----|-----| | OpenAI Codex | codex exec --full-auto | mcp__codex__codex |

More agents joining the crew soon. PRs welcome — bring your own worker.

Examples

See examples/:

• single-task.md — One agent, one assignment
• parallel-tasks.md — Three agents deployed simultaneously
• background-mode.md — Agents in the back room

FAQ

Q: Why not just use Claude Code for everything? A: You could. But a good boss delegates. When you have 4 independent tasks, deploying 4 agents to work in parallel is 4x faster. Claude stays focused on strategy and quality.

Q: What if I don't have tmux? A: Agents work in the back room (background processes). Claude checks on them via log files. Not as pretty, but gets the job done.

Q: Can I use agents other than Codex? A: Yes. The skill is agent-agnostic. Any CLI tool that accepts a prompt via stdin can be added to the roster.

Q: What if an agent hangs waiting for input? A: Every assignment starts with "You have one job. No questions. No brainstorming." Agents that try to chat instead of code get that beaten out of them by the preamble.

Q: What if two agents edit the same file? A: Claude detects conflicts during inspection, resolves them, and re-runs verification. It's the boss's job to coordinate.

Contributing

PRs welcome. Especially:

• New agent backends (Gemini CLI, Kimi, Cursor, etc.)
• Task dependency graphs (do B after A clocks out)
• Better tmux layouts for large crews

License

MIT