AI writes code fast, but it's missing structure. So I built a methodology layer.

Ionut Lomer — Thu, 05 Mar 2026 15:15:05 +0000

Every time I start a new project with Claude Code, I'm doing the same thing: the first hour or so is just setting up the structure that should already exist. Writing CLAUDE.md, defining agents, roles, conventions... then if you finish and decide to make a new project, all over again.

The AI can write code fast. But I feel it's missing structure. It writes code in circles. No clear tasks, no definition of done, no ownership boundaries. You end up being the project manager, the QA, and the developer, except now you're also managing an AI that doesn't remember what it did last session.

The problem

When you work with AI agents, there's no shared methodology. Every session starts cold. The agent doesn't know what was decided yesterday, what's in progress, or what "done" means. You spend more time managing the AI than building your product.

A lightweight backbone

I think agentic AI needs a methodology layer. Not a heavy framework, just enough structure so every session starts with shared context and clear boundaries.

I started playing with this idea and built a CLI called Meto. Simple approach with token optimization in mind: you answer a few questions about your project and it generates everything you need to start working with AI agents immediately.

npx meto-cli init

What it generates

CLAUDE.md pre-filled with your vision, stack, and conventions
3 agent roles for planning, building, and validation
Kanban board as markdown files: backlog → todo → in-progress → testing → done
Product context with vision, tech stack, decisions, definition of done
Starter epics tailored to your stack selection
Agent Teams ready pre-configured for Claude Code's multi-agent feature

The methodology

The idea is simple: PM plans, developer builds, tester validates.

Each agent has clear file ownership boundaries. The PM never touches /src/. The developer never touches the backlog. The tester never fixes bugs, only flags them.

The kanban board is all markdown files. No external tools needed. The AI reads and updates them directly.

Agent Teams support

With Claude Code's new Agent Teams feature, you can run all three agents in parallel. Just start Claude in your project and say:

"Create an agent team with @meto-pm for planning, @meto-developer for building, @meto-tester for validation"

The scaffold comes pre-configured. .claude/settings.json enables Agent Teams, and each agent knows its boundaries so they don't step on each other's files.

Stack presets

The CLI tailors the scaffold to your tech stack:

Next.js + Supabase with Supabase CLI detection
React Native with Expo-focused starter epics
Node.js CLI with entry point, help, version, npm publish
Custom describe your own stack

The whole point

The AI should follow a methodology from the first session, not the tenth.

Open source, MIT licensed:

iLomer / Metho_agentic

Scaffold AI-powered projects with built-in methodology. Agents, memory, kanban and governance out of the box. Works with Claude Code.

Meto

Lovable gives you an app. Meto gives you a project -- built the right way, your way.

Meto scaffolds structured software projects with built-in methodology. You describe what you want to build, and Meto bootstraps a project with AI-generated epics, a sliced backlog, agent definitions, product context, and coding conventions -- ready for your first Claude Code session.

Quick Start

npx meto-cli init

Answer a few questions, and Meto generates a fully structured repository in seconds -- with AI-powered content if Claude Code is installed.

Already have a project? Run the audit to bring it up to standard:

cd your-existing-project
npx meto-cli audit

How It Works

Meto detects whether Claude Code is installed on your machine and offers two paths:

With Claude Code (AI-powered):

You answer 5 questions -- project name, description, target users, tech stack, output directory
Choose your workflow mode -- Sprint (sequential) or Swarm (parallel epic…

View on GitHub

npm: npmjs.com/package/meto-cli

Does anyone else feel this gap? How do you structure your AI-assisted projects?

DEV Community: Ionut Lomer