Aha Loop

English | 中文

From a single paragraph to a complete project — fully autonomous AI development

Aha Loop is a fully autonomous AI development system, extended from Ralph. It's not just an execution engine, but a complete AI development framework with planning capabilities, research abilities, and oversight mechanisms.

One paragraph description → Complete runnable project

Want to know the story behind Aha Loop? Read The Design Story

What Aha Loop Adds to Ralph

Original Ralph	Aha Loop Extensions
Starts from PRD	Added Vision → Architecture → Roadmap pre-phases
Direct Story execution	Added five-phase workflow: Research → Exploration → Plan Review → Implement → Quality Review
Single execution path	Added autonomous parallel exploration (AI decides when to explore, creates worktrees, evaluates)
No oversight	Added God Committee independent supervision
Result-oriented	Added complete observability logging

If Ralph is an execution engine, Aha Loop adds the brain (planning layer) and eyes (oversight layer).

System Architecture

User: "I want to build an AI gateway..."
│
▼
┌───────────────┐
│ Vision Builder │ Interactive Q&A, build vision ← Aha Loop
└───────────────┘
│
▼
┌───────────────┐
│ Vision Skill │ Analyze vision, extract requirements ← Aha Loop
└───────────────┘
│
▼
┌───────────────┐
│ Architect │ Research tech, select architecture ← Aha Loop
└───────────────┘
│
▼
┌───────────────┐
│ Roadmap │ Plan milestones and PRDs ← Aha Loop
└───────────────┘
│
▼
┌────────────────────┐
│ PRD Execution Loop │
│ ┌──────────────┐ │
│ │ Research │ │ Research technology ← Aha Loop
│ │ Plan Review │ │ Review plan ← Aha Loop
│ │ Implement │ │ Write code (Ralph core)
│ │ Quality Check│ │ Quality review
│ └──────────────┘ │
└────────────────────┘
│
God Committee monitors throughout ← Aha Loop
│
▼
PROJECT COMPLETE

Core Principles

Principle	Meaning
Unlimited Resources	No limits on AI's compute, network, storage
Never Give Up	Auto-retry, auto-repair on issues
Research First	Learn before doing
Parallel Paths	Try all options, choose the best
Independent Oversight	God Committee monitors everything
Full Transparency	All decisions are traceable

Quick Start

Aha Loop is designed to work with Claude Code. Each phase is implemented as a Skill that you can invoke directly.

Workspace Mode (Recommended for External Projects)

Aha Loop can operate on any project without copying files into the Aha Loop repository. All Aha Loop data is stored in .aha-loop/ within your project:

# Initialize workspace in your project
/path/to/aha-loop/scripts/aha-loop/orchestrator.sh --init-workspace /path/to/your-project
# Run Aha Loop on your project
/path/to/aha-loop/scripts/aha-loop/orchestrator.sh --workspace /path/to/your-project --build-vision
# Or set environment variableexport AHA_LOOP_WORKSPACE=/path/to/your-project
/path/to/aha-loop/scripts/aha-loop/orchestrator.sh --build-vision

The --init-workspace command will:

• Create .aha-loop/ directory with all necessary subdirectories
• Copy skills to .claude/skills/ (required for AI agents)
• Copy templates, AI instructions, and example files

Windows Setup

If you're on Windows and symlinks aren't working (skills directory shows as text file), run:

git config core.symlinks true
Remove-Item".claude\skills"-Force; cmd /c mklink /D ".claude\skills""..\.agents\skills"Remove-Item".codex\skills"-Force; cmd /c mklink /D ".codex\skills""..\.agents\skills"

Phase 1-3: Planning (Manual with Claude Code)

Run Claude Code in your project directory, then use skills via /skill-name your description:

# Step 1: Build Vision - Interactive Q&A to construct project vision
/vision-builder I want to build an AI gateway that unifies multiple LLM APIs
# Step 2: Analyze Vision - Extract structured requirements
/vision Analyze my project vision
# Step 3: Design Architecture - Research tech and select architecture
/architect Design the system architecture
# Step 4: Create Roadmap - Plan milestones and PRDs
/roadmap Create project roadmap and break down into PRDs

Each skill generates corresponding documents:

• project.vision.md - Your project vision
• project.vision-analysis.md - Structured requirements
• project.architecture.md - Technical architecture
• project.roadmap.json - Milestones and PRD queue

Phase 4: Execution (Autonomous)

Once planning is complete, run the autonomous execution loop:

# Run PRD execution loop
./scripts/aha-loop/orchestrator.sh

The system will automatically execute a five-phase workflow for each Story:

1. Research: Fetch library source code, study implementations, generate research reports
2. Parallel Exploration: When AI identifies major technical decisions (multiple valid approaches), it autonomously:
   • Creates git worktrees for each approach
   • Runs parallel AI agents to implement each option
   • Evaluates results and recommends the best solution
3. Plan Review: Evaluate research and exploration findings, adjust plan if needed
4. Implement: Write code following research recommendations and exploration results
5. Quality Check: Validate implementation meets acceptance criteria

AI decides when each phase is needed. All of these are autonomous decisions. No manual intervention required.

Directory Structure

Workspace Mode (External Projects)

When using --init-workspace, your project will have:

your-project/
├── .claude/
│ └── skills/ # AI skills (copied from Aha Loop)
├── .aha-loop/ # Aha Loop working directory
│ ├── config.json # Execution config
│ ├── progress.txt # Progress log
│ ├── CLAUDE.md # AI instructions
│ ├── prompt.md # AI prompt
│ ├── templates/ # Document templates
│ ├── project.vision.md # Project vision
│ ├── project.roadmap.json # Milestones and PRDs
│ ├── research/ # Research reports
│ ├── exploration/ # Exploration results
│ ├── tasks/ # PRD documents
│ ├── knowledge/ # Knowledge base
│ ├── logs/ # AI thought logs
│ └── .god/ # God Committee
└── src/ # Your code (untouched)

Standalone Mode (Aha Loop Repository)

Aha Loop Repository:

AhaLoop/
├── .agents/skills/ # AI skill library (shared)
│ ├── vision-builder/ # Interactive vision building
│ ├── vision/ # Vision analysis
│ ├── architect/ # Architecture design
│ ├── roadmap/ # Roadmap planning
│ ├── research/ # Deep research
│ ├── parallel-explore/ # Parallel exploration
│ └── ... # Other skills
├── .claude/skills/ # Symlink → .agents/skills
├── .codex/skills/ # Symlink → .agents/skills
├── .god/ # God Committee
│ ├── config.json # Committee configuration
│ ├── council/ # Council chamber
│ ├── members/ # Member status (alpha/beta/gamma)
│ └── powers/ # Power records
├── scripts/aha-loop/ # Execution scripts
│ ├── aha-loop.sh # PRD executor
│ ├── parallel-explorer.sh # Parallel exploration
│ ├── config.json # Execution config
│ └── templates/ # Document templates
├── scripts/god/ # God Committee scripts
├── knowledge/ # Knowledge base
└── tasks/ # PRD documents

After using Aha Loop, your project will have:

your-project/
├── project.vision.md # Project vision (/vision-builder generates)
├── project.vision-analysis.md # Vision analysis (/vision generates)
├── project.architecture.md # Tech architecture (/architect generates)
├── project.roadmap.json # Milestones and PRDs (/roadmap generates)
├── tasks/ # PRD documents (auto-generated)
│ └── prd-001-xxx.md
└── logs/ # AI thought logs
└── ai-thoughts.md

Credits

• Ralph - Core PRD execution loop concept
• Claude - AI capabilities

License

MIT

Star History

---

Links:

• Example: AI-Gateway - API gateway built with early experimental Aha Loop
• Reference: snarktank/ralph

If you find Aha Loop interesting, please Star