Skip to content
Back to skills

Ralph Loop Edumesones Azure Demo

ASecurity

Autonomous feature development loop. Executes complete 9-phase cycle (Interview -> Think Critically -> Plan -> Branch -> Implement -> Verify -> PR -> Merge -> Wrap-Up) with minimal human intervention. Triggers on "/ralph", "start autonomous loop", "run ralph loop".

  • 666 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added June 10, 2026
ai-agentspythonbashgitfrontend

Works with

  • cli

Security analysis

A100/100

Scanned June 10, 2026

npx -y skills add majiayu000/claude-skill-registry --skill ralph-loop-edumesones-azure-demo --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ralph Loop Edumesones Azure Demo?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ralph Loop Edumesones Azure Demo
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/majiayu000-ralph-loop-edumesones-azure-demo/badge)](https://www.skillsdirectory.com/skills/majiayu000-ralph-loop-edumesones-azure-demo)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ralph-loop
description: Autonomous feature development loop. Executes complete 9-phase cycle (Interview -> Think Critically -> Plan -> Branch -> Implement -> Verify -> PR -> Merge -> Wrap-Up) with minimal human intervention. Triggers on "/ralph", "start autonomous loop", "run ralph loop".
globs: ["docs/features/**", "ralph-*.sh", "feature-loop-state.json", "activity.md"]
---

# Ralph Loop - Autonomous Feature Development

Executes the complete Feature Development Cycle (9 phases) autonomously, only pausing for:
- Human input during Interview (if spec.md is incomplete)
- Think Critically: red flags or low confidence level
- Merge approval (PR review)
- 3 consecutive failures

## Triggers

- `/ralph orchestrator [max_parallel]` - Start multi-feature orchestration
- `/ralph feature FEAT-XXX [max_iterations]` - Single feature loop
- `/ralph status` - Check running loops
- `/ralph stop` - Stop all loops
- "Start autonomous loop"
- "Run ralph on FEAT-XXX"

## Architecture

```
RALPH LOOP ARCHITECTURE

ORCHESTRATOR (ralph-orchestrator.sh)
- Reads docs/features/_index.md for Pending features
- Creates git worktrees for isolation
- Launches feature loops in parallel (max N)
- Monitors for merge completion
- Cleans up worktrees after merge

        |
        +--> Worktree: ../project-FEAT-001-loop/
        |    --> ralph-feature.sh FEAT-001
        |
        +--> Worktree: ../project-FEAT-002-loop/
        |    --> ralph-feature.sh FEAT-002
        |
        +--> Worktree: ../project-FEAT-003-loop/
             --> ralph-feature.sh FEAT-003
```

## Feature Loop Phases

```
FEATURE LOOP - 9 PHASES

Iteration 1: INTERVIEW
- Read spec.md
- If Technical Decisions has TBD -> PAUSE for human input
- If complete -> emit INTERVIEW_COMPLETE

Iteration 2: THINK CRITICALLY (11-Step Protocol)
- Determine depth: full/medium/light/skip
- Execute analysis protocol -> analysis.md
- If red flags or low confidence -> PAUSE for review
- If safe -> emit ANALYSIS_COMPLETE

Iteration 3: PLAN
- Read spec.md + analysis.md (BOTH required)
- Generate design.md with architecture
- Generate tasks.md with ordered checklist
- emit PLAN_COMPLETE

Iteration 4: BRANCH
- git checkout -b feature/XXX-name
- emit BRANCH_COMPLETE

Iterations 5-N: IMPLEMENT
- Process 3 tasks per iteration
- Mark [in progress] -> implement -> mark [x] -> commit
- emit IMPLEMENT_PROGRESS until all done
- emit IMPLEMENT_COMPLETE when finished

Iteration N+1: VERIFY (if frontend changes)
- Run browser E2E tests
- emit VERIFY_COMPLETE

Iteration N+2: PR
- git push origin feature/XXX
- gh pr create
- emit PR_COMPLETE

Iteration N+3: MERGE
- Poll gh pr view for merge status
- emit MERGE_WAITING until approved
- emit MERGE_COMPLETE when merged

Iteration N+4: WRAP-UP
- Create context/wrap_up.md
- Document metrics and learnings
- emit WRAPUP_COMPLETE
- emit FEATURE_COMPLETE
```

## Pause Conditions

The loop pauses (doesn't terminate) when:

1. **Human Input Needed** (return code 3)
   - spec.md has TBD values that can't be auto-filled
   - User must complete interview manually

2. **Think Critically Requires Review** (return code 3)
   - Step 2: Assumption with Low confidence + High impact
   - Step 9: Critical red flag identified
   - Step 11: Overall confidence = "Low"
   - User must review analysis.md and decide how to proceed

3. **Waiting for External Action** (return code 2)
   - PR created, waiting for review approval
   - Polls every 60 seconds

4. **Too Many Failures** (3 consecutive)
   - Something is broken
   - User must investigate and fix

## Integration with Feature Cycle

Ralph Loop fully integrates with existing Feature Development Cycle (9 phases):

| Existing Command | Ralph Equivalent |
|-----------------|------------------|
| `/interview FEAT-XXX` | Auto-executed if spec needs completion |
| `/think-critically FEAT-XXX` | Auto-executed with depth based on complexity |
| `/plan FEAT-XXX` | Auto-generated from spec + analysis |
| `/git "msg"` | Auto-commit after each task |
| `/git pr` | Auto-created after implementation |
| `/wrap-up FEAT-XXX` | Auto-generated after merge |

## Usage Examples

```bash
# Start orchestrator for all pending features (max 3 parallel)
./ralph-orchestrator.sh 3

# Run single feature loop
./ralph-feature.sh FEAT-001-auth 15

# Check status
./ralph-orchestrator.sh --status

# Stop all loops
./ralph-orchestrator.sh --stop
```

## Prerequisites

- `git` and `gh` (GitHub CLI) installed and authenticated
- `claude` CLI available
- `python3` for JSON manipulation
- Features defined in `docs/features/_index.md`

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.