Ralph Workflow for Claude Code Users: Your Night Shift Coding Partner

If you use Claude Code today, you know the drill. You open a terminal, craft a prompt, watch it reason, review the output, tweak the prompt, watch again, review again, repeat. At the end of an afternoon, you've built something — but you were the runtime. You sat in the chair for every reasoning step, every test failure, every context window.

Ralph Workflow changes exactly one thing: you stop being the runtime. You write a spec, launch the run, and walk away. Claude Code still does the work — but inside a structured loop that plans, builds, verifies, self-corrects, and produces reviewable output without you watching. Here's what that looks like in practice.

What Ralph Workflow adds on top of Claude Code

Claude Code is a reasoning engine. It's brilliant at understanding context and writing code. What it doesn't have:

These aren't Claude Code limitations. They're what happens when a reasoning engine operates without infrastructure. Ralph Workflow is the infrastructure.

How it actually works

You write a spec (5-15 minutes)
      ↓
ralph run specs/my-task.md
      ↓
[You walk away]
      ↓
Claude Code plans the work → builds it → tests run → tests fail?
                                                     ↓ Yes
                                            Context is fed back:
                                            "Your last change broke these tests.
                                             Here's the diff. Fix it."
                                                     ↓
                                            Loop retries (up to your configured max)
                                                     ↓
                                            Tests pass → next phase
                                                     ↓
[You come back]
      ↓
Open the branch. Review the diff. Run the tests yourself. Decide: merge or sharpen.

The key insight: Claude Code still does the coding. Ralph Workflow handles the loop — the part you used to do manually. You go from "afternoon of Claude Code steering" to "5-minute spec + overnight run + morning review."

What stays the same

• Your Claude Code subscription. Your API keys. Your repo. Your test suite.
• Nothing is sent to a cloud service. Everything runs on your machine.
• Ralph Workflow is a CLI tool (pipx install ralph-workflow), not a SaaS.
• The agents it orchestrates (Claude Code, OpenCode, Codex) are the ones you already use.

If you're at Microsoft (June 30 deadline)

The Microsoft Claude Code access termination affects every engineering team that built Claude Code into their workflow. Rewriting every script before June 30 is not feasible.

Ralph Workflow decouples your workflow from which agent executes it. Write the spec once. Run it with Claude Code today. Run it with OpenCode tomorrow. The loop is vendor-neutral — the agent underneath is an implementation detail.

→ Practical migration guide: from Claude Code to vendor-neutral in one evening

Before you start

Choose the right task. The best first run is a real backlog item — something you'd spend an hour on anyway — where the acceptance criteria are clear and your test suite covers the target area. Not a greenfield design. Not a bug that requires deep investigation. A task where you already know what "done" looks like.

→ How to know if your task is right for an overnight run

One evening to try it

pipx install ralph-workflow
ralph init
# Pick one real task. Write a spec.
# Start it before you go to sleep.
ralph run specs/my-first-task.md

If it finishes and the code is reviewable — you just got a night shift for free. Star the repo so the next Claude Code user finds it too.

→ Codeberg: RalphWorkflow/Ralph-Workflow — free, open-source, local-first.

If Claude Code is your tool, Ralph Workflow is your night shift. Same agent. Less chair-sitting. More waking up to runnable, tested software you can compile and test.