kdd-book reads a book or a technical guide and sorts every piece of advice into three honest buckets: what a computer can verify automatically, what is a real technique that still needs human judgment, and what is just background knowledge. For every piece that's verifiable, it then builds an automatic checker and a hands-on exercise with a solution.
A best-practices book tells you what to do. But nobody tells you, while you're writing code, whether you're actually following it — not until someone else reviews your work, or the bug is already in production.
"Name your variables clearly." "Don't repeat yourself." These are correct ideas, but no computer understands them as written.
Some advice has a clear, checkable rule behind it. Some depends on context and a developer's judgment. Treating them the same is the most common mistake.
Without an automatic check, "practicing" a book is just re-reading it. kdd-book builds the exercise and the checker for every technique that's genuinely testable.
The same method has been applied, so far, to 15 different sources — from style guides to technical specs to purely philosophical texts.
Every idea in the book becomes a knowledge node: a title, a description, and where it came from. Nothing is invented — everything traces back to the original text.
Every node gets one of three labels: measurable (there's a clear rule), judgment (it's real, but has no binary threshold), or knowledge (context, not a technique).
For every measurable technique, an instrument checks it in seconds, plus an exercise with a broken version and a correct one — proving the instrument tells them apart.
One real example of each category, taken from the sources already processed.
PEP 8, Python's style guide. An exact rule: it can be counted and checked in any file.
Clean Code, by Robert C. Martin. A real, valuable piece of advice — but there's no threshold separating a good name from a bad one.
The Zen of Python. It's design wisdom, not a technique with a property to check.
Neither source below "failed." Each one measured exactly what there was to measure.
A style guide designed to be enforced by a tool. Almost everything it says has an exact rule behind it.
19 lines of design wisdom ("simple is better than complex"). None of them has a binary threshold — and forcing one would misrepresent what the text actually says.
Several of the newest sources were built by an external AI agent, unsupervised during triage. Every result was re-run from scratch, every rule was deliberately sabotaged to confirm it genuinely catches the mistake it claims to catch, and every forbidden shortcut was manually reconstructed to prove the checker blocks it. Across two separate rounds, that verification found real defects — once, incomplete coverage; once, a test that didn't actually block the shortcut it claimed to block — and both were fixed before the result was accepted. No claim is accepted just because the agent said it worked.