Paper to Code

Turn a paper with no released code into the smallest implementation that can answer one question: does the method work as the paper claims?

The output is a verdict, not a codebase. Everything in this skill serves fast falsification: if the method doesn't work, you want to find out in an afternoon at toy scale, not after a week of building infrastructure around a broken mechanism.

When not to use

• The paper has official code → read/audit that instead (use reproducibility-audit if the question is whether the released artifact reproduces the results).
• The user wants to adapt a paper's component into their own codebase or design a full training setup → pytorch-training-recipe.
• The user wants to understand the method, not run it → professor-mentor-technical-teaching or flow-deep-understanding.

Process

1. Reduce the paper to one testable claim

Before writing anything, identify the single mechanism the reproduction should test. Papers bundle many things (new component + tuned baseline + data tricks); reproducing all of it is expensive and unnecessary for a sanity check.

State the claim in one falsifiable sentence, e.g. "Replacing softmax attention with mechanism X preserves accuracy while reducing memory" — not "reproduce Table 2". Confirm this framing with the user if the paper makes several claims.