HomeProjectsLearnVision AboutJournalContact

Design. Build. Evolve.

← Learn hub

Learn · The Builder's Path

From vibe coding to real engineering

Anyone can prompt an AI into a demo that works once. This is the interactive path to building tools you can actually trust — and use to push your own work in mechanical engineering, robotics, and physical AI.

Your progress 0%

"If it was as easy as everyone says it is on social media, we'd all be done already."

The Iceberg

Vibe coding is the tip. Engineering is everything beneath.

The shiny part — chatting with an AI and watching code appear — is real and powerful. But durable tools live below the waterline. Tap any layer to see what it means and why it matters.

Above the water · the easy 10%

Vibe Coding

Describe what you want in plain language and watch code appear. This is the genuine superpower — it collapses the time from idea to first draft. Tools like Claude Code make it conversational.
A script that runs once on your machine. Perfect for exploring an idea — but a demo is not yet a tool other people (or robots) can depend on.
Fast and satisfying. The trap is that you don't yet understand why it works — so when it breaks, you're stuck. This path closes that gap.
Below the water · the real 90%

Production Reality

Git, branches, pull requests. How you experiment safely, undo mistakes, and let an AI edit your files without fear. Covered in Track 01.
How you prove code works and keep it working — automatically, on every change. The difference between "ran once" and "reliable." See Track 03.
Docker, dependencies, lockfiles, servers, edge devices. Making your tool run the same way everywhere — including on real hardware.
Storing sensor streams, talking to other systems, validating inputs you don't control. Most engineering is, quietly, data handling.
Knowing what to build and which tradeoffs to make. The agent accelerates the typing; this path builds the judgement that directs it.

Good news: with Claude Code as your partner, you don't have to memorise the depths. You learn to recognise them, ask the right questions, and let the agent do the heavy lifting — under your direction.

How this works

Four tracks. Learn the concept, then build with it.

Each lesson pairs a short "why it matters" with a concrete command to run in Claude Code right away. Mark lessons complete as you go — your progress is saved in this browser. Then practise live in the Command Dojo and Prompt Lab below.

Specialised module

Robotics Study Plan

A full personalized robotics path — self-assessment with gap analysis, a competency curriculum with a study-time calculator, hands-on projects, and a 24-month roadmap.

Open module
Command Dojo

Try the workflow in a safe sandbox

A simulated Claude Code terminal. Type a command or tap one from the palette to see what it does — no risk, no setup. This is muscle memory for the real thing.

claude-code — practice session
Welcome to the Command Dojo.
This is a practice terminal — nothing here touches a real project.
Type help and press Enter, or tap a command on the right.
 
❯

Essential commands

The vocabulary that separates supervising an agent from just chatting with it.

Prompt Lab

Your output is only as good as your ask

The highest-leverage skill in AI-assisted engineering isn't typing — it's specifying. Write a prompt below and get instant feedback on what would make it sharper.

Grade my prompt

Write the instruction you'd give Claude Code, then check it against five marks of a strong prompt.

The same request, two ways

Notice how the strong version supplies context, constraints, and a definition of done.

✕ Vague "make a script to read my sensor"
✓ Specific "Write a Python script that reads my IMU over serial at 100 Hz, applies a low-pass filter, logs to CSV with timestamps, and handles disconnects gracefully. Use only pyserial + numpy. Show me a plan first."

Context · Constraints · A "done" condition · Plan-first. Master this and the agent feels twice as capable.

Build for your domain

Guided first projects in your field

Skills stick when they're applied to real work. Each of these is a small, finishable build — copy the starting prompt straight into Claude Code and go.

Mechanical Engineering

Parametric part generator

Turn a repetitive CAD task into code: a script that produces a configurable part and exports it for manufacturing.

  • 1Plan the parameters with the agent
  • 2Generate geometry with CadQuery
  • 3Export STL + a quick render
  • 4Add tests for valid dimensions
Starting promptUsing CadQuery, build a parametric mounting bracket with configurable hole spacing and thickness. Plan first, then implement, then export an STL.
Robotics

Sense-and-react node

A small ROS 2 node that closes the loop: read a sensor, decide, and command motion — tested in simulation first.

  • 1Scaffold a ROS 2 Python node
  • 2Subscribe to a laser scan
  • 3Publish a velocity command
  • 4Validate in Gazebo before hardware
Starting promptScaffold a ROS 2 node that stops the robot when an obstacle is within 0.5 m of the laser scan. Comment every line and tell me how to run it in simulation.
Physical AI

Sensor-fusion mini-pipeline

The data backbone of physical AI: ingest noisy signals, filter them, and produce a clean stream a model can learn from.

  • 1Generate synthetic noisy data
  • 2Implement a Kalman filter
  • 3Log clean output for training
  • 4Plot raw vs filtered to verify
Starting promptBuild a 1-D Kalman filter for a noisy distance sensor, generate synthetic data, log the result, and plot raw vs filtered. Explain each parameter.
Knowledge check

Quick gut-check

Three questions on the mindset that turns an AI assistant into an engineering multiplier. Pick an answer to reveal the reasoning.

Question 01

You're about to ask Claude Code to add a non-trivial feature. What's the highest-leverage first move?

Planning before building is the core discipline that separates engineering from vibe coding. Agreeing on the approach while changes are still cheap prevents wasted work and surfaces risks early.
Question 02

Why keep a CLAUDE.md file in your project?

CLAUDE.md is durable memory. It means you don't re-explain your project every session, and the agent makes choices that match your conventions. Generate a first draft with /init.
Question 03

Your AI-written script worked once. What makes it a tool you can actually trust?

"Ran once" is the tip of the iceberg. Tests prove it keeps working, a declared environment makes it run anywhere, and version control lets you change it safely — that's the production reality below the waterline.
Keep building

The tools you need don't exist yet. Build them.