SAGA — A Claude Code Plugin

Structured Agile Governance Agent

---

Who's Behind This

As an indie hacker, I kept starting projects, getting lost in the code, and forgetting which feature was supposed to do what. Needed a way to ship faster without losing track.

Found Geoffrey Huntley's Ralph Wiggum loop and Ryan Carson's workflow. Game changer. But I wanted something I could just run inside Claude Code without copying bash scripts around.

So I made this. It's not a replacement for Ryan's ralph — that works great. This is just how I like to work: everything inside Claude Code, with some extra tracking so I don't lose my mind on bigger projects.

---

What's This?

SAGA runs the Ralph loop as a Claude Code plugin. No terminal scripts. No jq. Just /saga commands.

I added SRS (structured requirements) instead of PRD because I kept forgetting what my own features were supposed to do. Now everything has an ID and I can actually trace what code does what.

The Ralph Thing

If you don't know Ralph, here's the gist from Geoffrey Huntley:

while :; do cat PROMPT.md | claude ; done

Run AI in a loop. Each iteration is fresh. Memory lives in files — git commits, progress logs. Define what "done" looks like, let it figure out how.

Geoffrey's insight: "LLMs are amplifiers of operator skill." When it fails, you don't blame the tool. You add better instructions. Failure is data.

Ryan Carson ships features while sleeping — three Ralph instances, three branches, three features. That's the dream.

How I Use It

My daily:

1. Morning — /saga init on a new idea, spend 20 mins on /saga spec really thinking through what I want
2. Let it run — /saga execute while I do other stuff
3. Check gaps — /saga trace to see what's covered, /saga gaps to see what's not
4. Tune — When something breaks, I add notes to the spec. That's the "signs" thing Geoffrey talks about.

For my simple projects, PRD would be fine. But I'm building stuff I want to maintain, maybe sell someday. So I use SRS — numbered requirements, traceable stories. When I come back in 2 weeks, I actually know what's going on.

Quick Start

/saga init        # setup
/saga spec        # think through requirements
/saga plan        # create stories
/saga execute     # run the loop

SAGA vs Ryan's Ralph

Not better. Different use case.

	ralph	SAGA
Setup	Copy scripts, install jq	/plugin install
Run	Terminal: ./ralph.sh	Inside Claude: /saga execute
Format	PRD	SRS
Best for	Quick features	Projects you'll maintain

Use Ryan's ralph if you like working in terminal. Use SAGA if you want everything inside Claude Code.

The Workflow

1. Spec first — The better your requirements, the better the output. I spend real time here.

2. Small stories — Each one fits in a context window. "Build auth" is too big. "Add email validation" is right.

3. Fast feedback — Tests matter. Without them, broken code compounds.

4. Tune with signs — It will fail. Add instructions. Try again. That's the process.

Commands

Command	What
/saga init	Setup
/saga spec	Requirements
/saga plan	Stories
/saga execute	Run loop
/saga status	Progress
/saga trace	Coverage
/saga gaps	What's missing

Install

/plugin marketplace add nooqta/saga
/plugin install saga@nooqta-saga

That's it.

Requirements

Just Claude Code CLI. That's it.

PM Sync

Optional GitHub/GitLab integration. Creates issues, closes them when done, makes PRs. Set up during /saga init if you want it.

Optional: Stitch AI

Building UI? Stitch AI adds design memory. I use it sometimes. Totally optional.

Credits

• Ralph Wiggum loop — Geoffrey Huntley
• ralph — Ryan Carson
• Ralf — my earlier attempt at this
• Built by Nooqta

License

MIT

---

Still figuring this out. If you try it and have ideas, open an issue.