The most fitting method would be to be to train an LLM on the Claude Code source-code (among other data).

Then use Anthropic's own argument that LLM output is original work and thus not subject to copyright.