Learn by shipping

The project ladder

Twelve builds that take you from your first deployed page to a spec-driven multi-agent app. Every rung has a real done — a live URL, passing tests, or a green eval. That’s the whole point: you don’t finish a lesson, you finish a thing.

Beginner Learn the loop and ship something live. No prior coding required.

Personal portfolio / landing page

The core loop: prompt → generate → preview → deploy, and reading diffs.

done → Live on a public URL (Vercel / Cloudflare Pages).

Any

CLI habit / to-do tracker

Letting the agent run and test its own code; reading errors back to it.

done → Runs in your terminal and persists between sessions.

Claude CodeCodex

"Explain this repo" exercise

Context referencing (@files) and codebase Q&A — not just generation.

done → A written architecture summary of a real open-source repo.

CursorClaude Code

Chrome extension / userscript

Working inside a platform’s constraints (manifests, permissions).

done → Loads in your browser and does one useful thing.

Any

Intermediate State, APIs, auth, data — real backends and your first tools.

Full-stack CRUD app with auth + DB

Schema design, env vars & secrets, a real backend.

done → Deployed, you can sign in and your data survives a refresh.

CursorClaude Code

API dashboard (weather / finance)

API keys, async, error & loading states, rate limits.

done → Live dashboard pulling real data, handling failure gracefully.

Any

Your first MCP server

How agents get tools — build one and test it with MCP Inspector.

done → Your editor’s agent can call your tool.

Claude CodeCursor

Custom slash command + subagent

Automating your own workflow; CLAUDE.md / AGENTS.md.

done → A /command or subagent that does a repeatable task for you.

Claude CodeCodex

Discord / Telegram bot with an LLM

Webhooks, long-running processes, deploying a service.

done → A bot live in a channel, responding.

Any

Advanced Agents, evals, and orchestration. Build like it’s going to production.

A research agent that does a real task

Tool use, loops, planning — search → summarize → deliver.

done → It completes the task end-to-end without you steering each step.

Claude CodeCodex

Eval suite + CI gate on an AI feature

Measuring AI output; failing the build when a prompt regresses.

done → A bad prompt turns your GitHub Actions check red.

Any

Spec-driven multi-agent capstone

Decomposition, parallel agents against a spec, verify-then-ship.

done → A real multi-feature product, built to acceptance criteria and live.

Claude CodeCodexSpec Kit

Free · the newsletter

Get the build logs + the Mirror System

One email a week: a real build broken down, plus a working reference agent you can clone. No fluff, unsubscribe anytime.