콘텐츠로 이동

AI-DLC for Claude Code

A structured, adaptive software development workflow that guides AI through disciplined three-phase development.

Get Started -- Just run /aidlc and the workflow adapts automatically.

View on GitHub -- Apache-2.0 licensed. Adapted from AWS Labs AI-DLC.

Origin

AI-DLC was originally developed by AWS Labs for Kiro. This plugin adapts the methodology for Claude Code with multi-agent delegation and model tiering.

What is AI-DLC?

AI-DLC (AI-Driven Development Life Cycle) is a methodology that brings structured software engineering discipline to AI-assisted development. Instead of ad-hoc coding, AI-DLC guides the AI through a three-phase lifecycle with human approval gates at every stage.

Why AI-DLC?

Without AI-DLC With AI-DLC
AI jumps straight to coding Structured analysis before implementation
No documentation trail Full audit trail with ISO 8601 timestamps
One-size-fits-all approach Adaptive depth based on complexity
Single model for everything Opus for strategy, Sonnet for execution
Hard to resume interrupted work Session continuity via state tracking
No approval gates Human-in-the-loop at every stage

Key Features

One Command

Just run /aidlc and the workflow adapts to your project automatically.

  • Adaptive Depth -- Simple bug fixes skip unnecessary stages. Complex systems get full architectural treatment.
  • Deep Questioning -- INCEPTION stages ask 15-30 practical questions across mandatory categories with multi-round follow-up. No more shallow Q&A.
  • Three-Phase Lifecycle -- INCEPTION (what and why) → CONSTRUCTION (how) → OPERATIONS (deploy and monitor).
  • Human-in-the-Loop -- Every stage requires your explicit approval before proceeding.
  • Multi-Agent Delegation -- Opus handles strategic reasoning, Sonnet handles volume work, Haiku handles fast detection.
  • Parallel Unit Execution -- For multi-unit projects (3+), independent units execute simultaneously in parallel groups, reducing total build time.
  • Full Audit Trail -- Every decision documented with ISO 8601 timestamps.
  • Stage Banners (MOTD) -- Every agent displays a formatted banner on start showing phase, stage number, agent name, model, and capabilities.
  • Session Continuity -- Interrupted workflows can be resumed from where you left off.
  • Brownfield Support -- Deep reverse engineering for existing codebases with 8 analysis artifacts.
  • PR Review -- Standalone utility for analyzing PR diffs across 6 categories (correctness, security, performance, consistency, testing, documentation).
  • CI Setup -- Standalone utility to generate CI/CD pipelines, PR review workflows, and issue/PR templates with automatic tech stack detection.
  • Dependency Graph -- Optional graph-based code dependency analysis with multi-backend support (File/Neo4j/Neptune), impact analysis, Mermaid visualization, PNG export, 9-point deployment verification, CGIG compilation repair, and change-aware test prioritization. E2E verified with Neo4j and Neptune backends.
  • GraphRAG -- Optional summary-based semantic code retrieval. Claude generates module summaries and community structure stored as graph properties -- no external embedding models required. Search by purpose, find related modules, understand code semantics.
  • CGIG Repair -- Compilation-Guided Iterative Graph-retrieval for automated error repair. 4 graph construction methods (Static/CGIG/Lightweight/Hybrid), 10 language-agnostic error categories, confidence-scored fix suggestions from dependency graph context.

The Three Phases

INCEPTION -- What and Why

Understands the problem space. Detects workspace type, gathers requirements, creates user stories, plans the execution, designs architecture, and decomposes into implementation units.

Workspace Detection → Reverse Engineering → Requirements Analysis
→ User Stories → Workflow Planning → Application Design → Units Generation

CONSTRUCTION -- How

Implements the system. For each unit (sequentially or in parallel): designs business logic, evaluates non-functional requirements, maps infrastructure, generates code. Finishes with build and test.

Per-Unit Loop (sequential or parallel):
  Functional Design → NFR Requirements → NFR Design
  → Infrastructure Design → Code Generation

Build and Test (after all units)

OPERATIONS -- Ship

Deployment checklist, CI/CD pipeline, Dockerfile, environment templates, developer README, and post-deployment verification scripts. Ensures the generated project is immediately deployable and developer-ready.

Adaptive Execution

Not every stage runs every time. The workflow adapts to your needs:

Condition What Happens
Simple bug fix Only essential stages: detection, requirements, planning, code gen, test
New feature (greenfield) Full INCEPTION + CONSTRUCTION treatment
Brownfield modification Adds reverse engineering, adapts scope to existing codebase
Infrastructure-only change Skips user stories and functional design

Note

You can override any recommendation at the Workflow Planning approval gate.

References

AI-DLC Methodology

Code Graph and CGIG Research

The CGIG feature (v1.8.0) is based on research from SELENE Lab, Korea University. These papers inform graph-based code generation and repair: