Mining 9 Months of AI Chat Logs into a Workflow Plugin

BiosRios · reddit · 2026-07-11

The author mined their own local AI chat logs—spanning **9 months, 1,656 sessions, and roughly 3 million tokens**—to create a plugin that helps agents better understand their working style. ### His Approach - Filtered out 95% of the noise, keeping only the actual inputs - Deduplicated the data while preserving patterns that recurred across multiple sessions - Extracted "decision criteria" from the history of various tools, rather than simply summarizing code ### What Was Mined: Work Habits, Not Code For example: - An instant aversion to generic UIs, preferring high-contrast, compact layouts - A preference for referencing first—often "copying something they like" rather than describing it verbally - Disliking being given 5 options; wanting a single, direct answer - A communication style that is short, direct, and without fluff ### The Author's Goal To let agents read these preferences, reducing default generalizations and aligning more closely with his judgment, thereby minimizing repetitive explanations. ### Important Notes - He emphasizes this isn't about having the model "summarize 9 months of logs" directly, but rather extracting characteristics through repeated cross-session evidence - He acknowledges that the current version consumes a lot of tokens initially, but plans to implement a more token-efficient incremental/caching approach later Finally, he shared the project link, encouraging everyone to try it out and provide feedback.

Original post →

More from coding & agent

coding & agent channel →