Vocabulary journal

The failure this fixes

You can follow a paper and still be unable to define its central term a week later. Reading produces recognition; code review and interviews require retrieval. A vocabulary journal is the smallest practice that converts one into the other.

One line, in your own words

For every term you meet, write one line that says what it is and what it is not. “Layout assignment chooses physical memory layouts and inserts copies where a consumer wants a different one; it is not fusion, which decides what runs in a single kernel.” The contrast clause does the work, because most confusion is between two neighbouring terms rather than about a term in isolation.

Write the line the day you meet the term. A definition copied from the paper is worthless, since the compression is the learning.

Retrieval, not review

Turn each line into a question and answer it from memory before you look. Re-reading a notebook feels productive and tests nothing. The prompts essay is the best short treatment of what makes a question answerable and specific; read it once, then rewrite your worst entries.

A model to imitate

The LLVM Lexicon is the genre at its best: one entry per term, each distinguishable from its neighbours, with the usual confusions named. Your journal should read like a smaller, personal version of it, restricted to terms you have actually used in code.

What this is not

Not a paper summary, not a glossary generated by copying, and not a substitute for the reading. If an entry could have been written without the source, it probably was.

Afterwards, for each theory node you have finished, you should hold three or four entries you can define cold. If you cannot write the entry, you did not read the node.

Things that break

Circular definitions, which are the default: “tiling is the process of tiling an operation”. They survive a spell check and fail the only test that matters, which is using the term in a sentence about your own code. The second failure is a journal that grows and is never queried, which is a second copy of the reading with extra steps.