Softwareaiclaudedebuggingclipythonopen-source
Nietzschian Debugger Runs on Claude and Refuses to Fix Your Bug, Asking Questions at Three Settings From Socrates to Zarathustra
A philosophy graduate turned developer in Lahore built a command-line debugger that only ever answers with a question, with a model-judge checking every reply so it never slips and hands over the fix.
The pitch on the README is one line: “It won’t fix your bug. It’ll fix you.” Dilawar Shafiq, a self-described “philosophy graduate, turned to software development” in Lahore, opens the project with a complaint about where coding has gone: “Copy-paste from ChatGPT, ship, repeat. Nobody understands their own code anymore.”
His answer installs with pip install nietzschian-debugger and runs as nietzschian debug "your bug". From there, “Every response is a question.” The tool is “built to withhold the answer, and a model-judge double-checks each reply to catch it when it slips,” which matters, because a chat model’s whole training pulls toward handing over the fix.
The intensity flag sets the tone. Socrates gives “Warm, guiding, scaffolded questions” for juniors. Nietzsche, the default, is “Direct, confrontational, no hand-holding.” Zarathustra is for “Seniors who want real pain.” It quotes philosophers to match the mood, “Nietzsche when you’re avoiding, Seneca when overwhelmed, Sun Tzu when you need strategy,” and it can read “YOUR actual code” when you name a file path, or your uncommitted changes with --diff.
There is a way out. Type I give up and you get “one lifeline question, then exit.” Type solved or found it and the session ends with a Growth Score charting how well you checked assumptions, gathered evidence and reached the root cause, saved locally so the trend shows over weeks. A nietzschian demo runs with no API key at all.
The project dates to February, when it was called Socratic Debugger and written in TypeScript before a rewrite in Python. It shipped to PyPI on October 27, went through five releases that day, and added OpenAI, Gemini, DeepSeek, Ollama and other back ends in v0.3.0. Shafiq’s own caveat sits in the docs: “the ‘never give answers’ discipline is sharpest on Claude. Smaller local models leak and over-ask more.” How long a developer under a deadline will put up with Zarathustra before typing I give up is the experiment every user runs on themselves.
The bet underneath tools like this is that AI coding help splits into two kinds, one that does the work and one that trains the person doing it, and that teams will want both. The first kind is winning on speed today. The second kind has a measurement problem the first doesn’t, since a developer who got better at debugging produces no metric a manager sees, which is why the Growth Score may turn out to be the most important feature in the repo.