Claude Code scripts: 50% one-off, not 69%

The author originally intended to publish an article arguing that 70% of AI-generated scripts are disposable and that this is fine. The number came from a single snapshot of Claude Code session logs, counting how many distinct sessions a script's name appeared in an execution command. That first count showed 69% of scripts had been used in only one session.

Three weeks later, before writing the article, the author reran the same count and got 50%. The reason was not that scripts had suddenly become more reusable, but that the earlier count included scripts created so recently that their second use had not yet arrived. Of the 156 scripts that were one-off on 09-08, 39 (25%) got a second use within the next three weeks, with a median gap of 12 days between first and second use. The body calls this right-censoring: events that have not happened yet get recorded as never happening.

The story also notes a shift in how scripts were being used. Scripts first run after 09-09 had a lower one-off rate for practical tasks (31%) than earlier ones (46%), which the author speculates reflects a move toward turning repeated procedures into reusable 'skills' rather than writing a fresh script each time. The scripts that survive longest are infrastructure (the run_jsx dispatchers) and routine business tasks called from those skills. The outcome hinges on whether the author's proposed fix — counting twice, separating recent scripts, and remeasuring before publishing — actually changes how these numbers are reported going forward.