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
Methodology agentic AI - CLI tool
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 a kanban board, agent definitions, product context, and coding conventions -- ready for your first Claude Code session.
Quick Start
npx meto-cli init
Answer a few questions about your project, and Meto generates a fully structured repository in seconds.
What Just Happened?
After running meto-cli init, your new project comes pre-loaded with everything you need to start building with discipline:
- CLAUDE.md -- a project instruction file that Claude Code reads every session, pre-filled with your vision, stack, and conventions
- Kanban board -- a task pipeline (backlog, todo, in-progress, testing, done) ready for your first sprint
- Agent definitions -- PM, developer, and tester agents configured to follow your methodology from day…
npm: npmjs.com/package/meto-cli
Does anyone else feel this gap? How do you structure your AI-assisted projects?
Top comments (0)