The Five Levels of
AI-Assisted Coding

You ask the AI a question. It answers it. What's the I2C address range for peripherals on an Arduino Uno? What does this linker error mean? What's the difference between SPI mode 0 and mode 3?

This maps directly onto what most people already do with search engines, except the AI synthesises and explains rather than returning links. For a maker digging into an unfamiliar datasheet, this is genuinely useful. It's also the level with the lowest ceiling — you're getting knowledge, not code.

You own the code. The AI fills in one piece of it. You're writing firmware for a sensor interface and you need a function that converts a 12-bit ADC reading to a voltage given a 3.3V reference.

You're still the architect. You decided the program structure, the data types, where this function sits. The AI handled one implementation detail. The key discipline is that you're not asking for the whole program — you're asking for a piece you can evaluate on its own terms.

The AI isn't filling in a snippet. It's implementing a complete, self-contained piece of your design. You need a Python class that manages a serial connection to an Arduino: opens the port, sends JSON commands, reads responses, handles timeouts, raises a clean exception if the connection drops.

You designed the interface; the AI wrote the implementation. You still need to understand the result well enough to test it and integrate it. Domain knowledge pays off here: knowing the edge cases — dropped connections, malformed responses — is what makes the specification precise enough for the AI to get it right.

You provide a specification. The AI returns a working file or sub-system. A concrete example: write a single-file Python script that reads sensor data from a serial port at 9600 baud, plots three channels live, and saves to CSV when the user closes the window.

You are the architect and the reviewer. You specify what the system does; the AI codes the whole thing. The quality of the output depends almost entirely on the quality of your specification. Underspecify and you get something that technically works but doesn't fit your requirements. This is where framing becomes the bottleneck.

The AI doesn't just generate code. It reads your files, runs commands, observes the output, and responds to what it finds. You set a goal; the AI executes the steps. You point Claude Code at a half-built serial plotter and say: "The graph freezes when the Arduino sends malformed data. Find the issue and fix it."

Claude Code reads the source files, identifies where the parser fails on unexpected input, writes a fix, runs the script to verify it, and reports what it changed. You didn't write any of that code. You reviewed the diff and decided whether to accept it. Level 5 is the most powerful mode — and the one that requires the most from you.