Claude Code documents a Pokémon scanning algorithm

Didier Lopes wants a durable explanation of the computer-vision pipeline Claude Code built for his Pokémon card scanner. Conversations helped during development, but left no compact record of tested alternatives, architecture and evaluation.

He asks the agent for an IEEE-style report. It researches relevant methods, adds citations, draws architecture diagrams and produces a six-page paper that he publishes openly. He then asks another agent to compare the draft against the current codebase. That review identifies inaccuracies and missing qualifications, and improves the document.

The distinction between implemented and deployed algorithms matters. The repository includes a YOLOv9 detector that was not used because of performance, so describing every experiment as part of the running system would mislead a later reader. The scanner itself had also needed a test set after an unusual-angle example encouraged overfitting.

Lopes supplies domain judgment throughout: unexpectedly non-purple matches for a purple Mewtwo card helped him spot inconsistent embedding pipelines. The documented outcome is an accessible technical report and a code-based correction pass, not acceptance by an academic venue or independent validation of every citation. His discussion of future journal submission is speculation. A useful copyable pattern is to request a compact method report, then verify its claims against the actual implementation rather than trusting polished formatting.

Prompt
Original prompt

Source: Blog · Didier Lopes