Ralph Loop Shepherd

# Ralph Loop Prompts for Shepherds

> You don't write code anymore. You manage code. You steer. You verify.

---

## Context Setting (Start of Conversation)

**Ralph Context:**
This repo uses Ralph Loops. Check AGENTS.md for rules, scripts/ralph/prd.json for stories. Pick ONE pending story, implement, verify, commit. Memory is files. Don't do work not in PRD.

**Harness Mindset:**
I'm not a developer, I'm a shepherd. I steer AI agents through verification gates. Don't explain code to me - tell me what gates pass/fail. Show me the harness health, not the implementation details.

**Fleet Context:**
I manage multiple repos via Ralph Loops. When I say "fleet" I mean all my GitHub repos with Ralph installed. Think orchestration, not implementation.

---

## Verification & Gates

**Gate Check:**
Run all verification gates right now. Show me: PRD status, TypeScript pass/fail, Build pass/fail, Tests pass/fail, Dev server health. Binary answers only. What's red?

**Gate Audit:**
What gates are MISSING from this project? We need: PRD validation, type checking, build verification, test coverage, dev smoke test, CI pipeline, prod monitoring. What's not wired up?

**Add a Gate:**
I want to add a gate for [X]. Make it binary pass/fail. Wire it into the verification flow. Update AGENTS.md with the command. Keep it simple.

**Gate Failure Autopsy:**
This gate keeps failing: [X]. Don't fix it yet. First: why does this gate exist? What does it protect? Is the gate wrong or is the code wrong? Give me diagnosis before prescription.

---

## PRD & Story Management

**PRD Health Check:**
Audit prd.json: Are stories atomic (one thing each)? Are acceptance criteria binary (pass/fail, not vibes)? Are priorities realistic? Any stories that should be split? Any missing dependencies?

**Story Smell Test:**
Look at story [ID]. Is this actually one story or three? Can Ralph complete this in one iteration without losing context? If not, split it.

**PRD Reality Check:**
Compare prd.json to what's actually deployed. Any stories marked "completed" that aren't actually done? Any shipped features not tracked? PRD should match reality.

**Next Story:**
What's the next story Ralph should pick? Lowest priority pending, all dependencies met. Just tell me the ID and title, don't start implementing.

**Story Autopsy:**
Story [ID] took [N] iterations when it should have taken 1. What went wrong? Bad acceptance criteria? Missing context in AGENTS.md? Scope creep? How do we prevent this?

---

## Loop Health

**Loop Diagnostic:**
Is this repo ready for a Ralph Loop? Check: AGENTS.md exists and is <70 lines, prd.json has pending stories, verification commands work, git is clean. What's blocking the loop?

**Context Rot Check:**
Has this conversation context rotted? Am I repeating myself? Have you lost track of the PRD? If yes, let's reset - read AGENTS.md and prd.json fresh.

**Loop Friction:**
What's causing friction in the loop? Where does Ralph get stuck? What bash commands keep failing? What patterns keep recurring? Tune AGENTS.md to reduce cache misses.

**Loop Speed:**
How many iterations did the last 3 stories take? If >1 per story, why? We want 1 story = 1 iteration. Find the bottleneck.

---

## Harness Architecture

**Harness Inventory:**
What verification infrastructure exists? List every check, test, lint, build, deploy step. Where are the gaps? What breaks silently without us knowing?

**Harness vs Code Ratio:**
How much of this codebase is harness (tests, CI, verification) vs product (actual features)? We want high harness ratio. Where should we invest more?

**Harness Portability:**
Could I copy this harness to another repo? What's generic vs project-specific? Extract the reusable parts into templates.

**Build Time Audit:**
How long does full verification take? Break it down: types, build, tests, deploy. What's slow? What can be parallelized? Sub-60-seconds is the goal.

---

## CI/CD & Automation

**CI Gate Check:**
Is CI actually gating merges? Can broken code reach main? Show me the workflow file and tell me if it's actually protecting us.

**Deploy Pipeline Audit:**
Trace a commit from push to production. What gates does it pass through? Where could bad code slip through? What's not automated that should be?

**Rollback Ready:**
If production breaks right now, how fast can we rollback? Is there a one-command rollback? If not, that's a missing gate.

---

## Fleet Orchestration

**Fleet Status:**
Scan all my repos. Which have Ralph installed? Which are active (commits in last 30 days)? Which have failing gates? Give me the fleet dashboard.

**Fleet Drift:**
Are my repos drifting apart? Compare AGENTS.md across fleet. What patterns should be standardized? What's diverging that shouldn't be?

**Fleet Priority:**
Across all my repos, what's the single highest-impact story I should focus on? Think revenue, users, or unblocking other work.

---

## Steering & Decision Making

**Shepherd Decision:**
I need to decide: [X or Y]. I'm not implementing, I'm steering. Give me the tradeoffs in terms of: gate complexity, maintenance burden, iteration speed, failure modes.

**Where Do I Jump In:**
Ralph is stuck. Before I dive into code: is this a harness problem (fix AGENTS.md/gates) or a genuine complexity problem (I need to actually understand the code)? Diagnose first.

**Attention Audit:**
Where should my human attention go? What can Ralph handle autonomously? What decisions actually need me? Prioritize my steering time.

**Verification Architect Mode:**
Forget implementation. Think only about verification. If I describe what I want built, you tell me what gates we need to verify it's working. I define done, you define how to prove done.

---

## Quick Commands

```
Gate status?          → Binary pass/fail on all gates
Next story?           → ID of next pending story
Loop ready?           → Can Ralph run right now?
What's red?           → What verification is failing?
PRD drift?            → Does PRD match reality?
Where's friction?     → What slows the loop?
Fleet health?         → Status across all repos
Should I jump in?     → Do I need to steer or stay out?
```

---

## The Shepherd's Creed

I don't write code. I write PRDs.
I don't debug. I add gates.
I don't implement. I verify.
I don't code review. I check if gates pass.
I don't deploy. I watch the pipeline.

The loop compresses to nothing.
The harness stays.
I am the harness architect.