Networking for Spies: Translating a Cyrillic Text with Claude Code (austinpatrick.substack.com)

An author used Anthropic’s Claude Code to build a simple, reproducible pipeline that translated a previously untranslated Russian spy handbook (Нетворкинг по‑русски) in minutes: they fetched the PDF from an open archive, extracted machine‑readable text deterministically with PyPDF2, split the corpus into small chunks, and fed those pieces to Claude Code (an interactive LLM terminal) to produce a reliable translation and reusable scripts. The workflow deliberately separates evidence collection (OCR/text extraction) from interpretation (LLM translation) to avoid adding variability or hallucination where deterministic tools are better suited. The code runs locally via Claude Code, avoiding extra API costs, and the author released the materials in a repository. For the AI/ML community this is a practical demonstration of how LLMs broaden access to non‑Anglophone primary sources and how proper tooling and mental models improve reproducibility and trust. Key technical takeaways: use deterministic libraries (PyPDF2) for extraction, chunk large texts to keep translation quality high, and keep an auditable intermediate step so you can verify LLM outputs against the original. The pattern reduces error, cost, and effort while highlighting risks (overloading models with raw PDFs invites poor translations) and opens opportunities for multilingual dataset creation, cross‑cultural research, and more rigorous provenance practices.