rabbit-workflow

Structured Claude Code workflow enforcing Machine First, Bounded Scope, and Designed Deprecation.

Policy is auto-injected every 20 prompts (configurable). Slash commands and skills included.

---

For developers (contributing to rabbit-workflow)

git

git clone https://github.com/USER/rabbit-workflow
git clone https://github.com/USER/rabbit-workflow my-name  # custom workspace name

---

For users (installing into your own workspace)

Your target workspace must not have an existing .claude/ directory.

git

git clone https://github.com/USER/rabbit-workflow
./rabbit-workflow/install.py /path/to/your/workspace   # explicit target
./rabbit-workflow/install.py                           # or install to $PWD

curl (installs to current directory)

cd /path/to/your/workspace
bash <(curl -fsSL https://raw.githubusercontent.com/USER/rabbit-workflow/main/install.py)

---

Commands

Command	Description
/rabbit-refresh	Manually re-inject policy
/rabbit-config prompt-threshold N	Set auto-refresh interval to N prompts (takes effect next session)
/rabbit-config prompt-threshold	Restore default auto-refresh interval
/rabbit-config allowed-tools [add|remove <tool>]	Manage Claude Code tool permissions
/rabbit-config bash-allow [add|remove <cmd>]	Manage Bash command permissions
/rabbit-config permissions [lock|unlock]	Lock/unlock archive/ and test/ owner write bit
/rabbit-config human-approval [true|false]	Manage Step 4 HUMAN-APPROVAL gate (true=gate ACTIVE/default, false=bypass ACTIVE/marker written)
/rabbit-project init <name>	Scaffold a new project directory

Configuration

Default refresh interval: 20 prompts. Change with /rabbit-config prompt-threshold N.

Applying rabbit anywhere

There is one work model. Every feature schema, every script, every subagent works the same regardless of where the feature directory lives: .claude/features/<x>/ (rabbit improving itself), projA/features/<y>/ (any project applying the rabbit discipline), or any other path.

The TDD subagent is dispatched with a SCOPE per invocation; the scope-guard hook enforces that scope. There is no rabbit-dev-mode vs user-mode dichotomy in the runtime.

# Scaffold a new feature anywhere (Python — no .sh scripts in rabbit)
python3 .claude/features/rabbit-feature/scripts/new-feature.py \
    projA/features auth-redirect --owner alice

# Audit every feature in a tree (moved to rabbit-feature-audit skill in
# rabbit-feature)
Skill("rabbit-feature-audit", args: "projA/features")

# File a bug (rabbit-file owns bug/backlog item lifecycle)
python3 .claude/features/rabbit-file/scripts/file-item.py \
    --feature auth-redirect --type bug \
    --title "login redirects loop on safari" --priority high

# Transition TDD state — same script for any feature dir
python3 .claude/features/tdd-subagent/scripts/tdd-step.py \
    transition projA/features/auth-redirect test-red

# Dispatch the TDD subagent onto a scope (sketched — typically the main session
# orchestrates this; the main session writes/removes the marker around the
# Agent call)
touch .rabbit-scope-active-auth-redirect
# (Agent dispatch with subagent_type: tdd-subagent, prompt with SCOPE: ...)
rm .rabbit-scope-active-auth-redirect

See .claude/features/rabbit-cage/docs/spec/spec.md for the rabbit-cage spec and .claude/features/tdd-subagent/docs/spec/spec.md for the TDD subagent protocol.

Uninstall

Remove .claude/ and CLAUDE.md from the installed workspace.