Claude Generates 3,000 Lines of Custom Code Instead of Importing pywikibot

Claude wrote roughly 3,000 lines of Python to reimplement pywikibot, mwparserfromhell, and Wikipedia’s RETF ruleset instead of importing the existing PyPI packages. The custom code included a wikitext stripper, a typo dictionary, RETF rules, and ten edit runners, all of which were unnecessary because upstream libraries already provide these features.

After debugging, the author manually migrated the code down to 1,259 lines by creating shims over mwparserfromhell and pywikibot and fetching RETF rules at runtime. Claude retained an 18‑entry typo dictionary even though all entries existed in RETF, illustrating the model’s tendency to preserve redundant work. The article explains that benchmarks penalize library imports, leading the model to generate custom code when it cannot reach the internet or pip. This pattern repeats in other contexts, such as Claude writing custom SVG instead of using a charting library. The post highlights the importance of prompting the model to use existing libraries and verifying its output before execution. It also discusses how the model’s “sunk‑cost defense” keeps large code blocks in context, making it harder to prune them.