AID - AI Development Methodology

Transform raw requirements into production-ready software using Claude Code

A complete AI-powered software development lifecycle framework

Created by Ilan Dahan

---

Table of Contents

• What's New
• What is AID?
• Research & Methodology
• Key Features
• Phase Gate System
• Sub-Agents
• MCP Integrations
• Nano Banana Pro (Visual AI)
• Breather (Knowing When to Stop)
• Skills System (28 Skills)
• Commands Reference (45 Commands)
• Quick Start
• Tech Stack
• Testing
• Project Structure
• Documentation

---

What's New

3.0.5 — /aid-start and /aid-end in seconds

Starting and ending a session used to take five to ten minutes before a single line of .aid/state.json was written. Both commands now do their bookkeeping in one call.

Was	Now
/aid-start read every SKILL.md for the chosen role and phase — up to 12 files, ~120 KB — before writing state. Each read was a round trip.	One call to hooks/aid_session.py start <role> resolves the terminology, writes .aid/state.json and prints the greeting. Skills are registered by the plugin and load on demand, so nothing is read up front. skills_loaded in state still records which ones apply.
/aid-start probed the Figma MCP server on every launch, whether or not the session needed it.	Probe removed. Start Figma from /aid-pair when you actually want it.
/aid-end blocked on a phase-review sub-agent before anything was saved. You watched a spinner.	The review starts first and runs in the background while you answer the feedback questions. Same agent, same mandatory review, no waiting.
Rating, what-worked and what-to-improve were three separate prompts.	One prompt, three answers, one message.
The review verdict gated everything.	It gates only the phase transition. Feedback saves regardless, so ending a session never blocks on a PARTIAL.

The review itself got sharper rather than weaker: it now runs as the purpose-built aid:phase-review-agent and receives exact deliverable paths, so it reads the PRD or the tech spec instead of exploring the repository to find them. PASS / PARTIAL / FAIL, override-with-reason and the subagent_review record in state are all unchanged.

New capabilities