Patched Claude instructions degraded a meeting-analysis workflow — errors doubled runtime

Psychobert · reddit · 2026-09-11

The author uses Claude projects to analyze large Plaud-transcribed meeting audio (2-6 speakers) and must produce structured minutes (work packages, issues, next actions) within 30 minutes of meeting close for audit and PMO. Over recent weeks output quality has collapsed: wrong stage gates, missed topics, ignored output format; runtime doubled from 5-7 minutes with results that can't be trusted. Pointing out errors only gets "you're right, I missed that." The author suspects ad-hoc instruction patches have bloated the project instructions and degraded output, and is weighing a clean rewrite — but doesn't know how to prevent drift. A meta point: this is AI patching a process problem created when the client eliminated the junior staff who used to do this well.

Original post →

More from coding & agent

coding & agent channel →