Every Friday night, I go back through this week's AI conversations
A developer has a fixed Friday ritual: run an analysis task, review this week's Claude and Codex sessions, distill reusable lessons into skills or scripts, and add missing context to notes. He says his role is becoming building environments and improving context.
A developer has set a fixed ritual for himself: every Friday night, he runs an analysis task and goes through this week's Claude and Codex sessions.
He is not just casually looking through chat logs. He handles them in three categories: reusable lessons get distilled into skills, tools, or scripts; missing context gets added to notes. The purpose is straightforward: make the agent execute more smoothly next time.
In his tweet, he wrote a self-deprecating line: my job is gradually turning into building environments and improving context.
That sentence is worth pausing on. People used to say an engineer's output is code; now part of that output has become things that pave the way for models: a clearly written background brief, a packaged script, a codified skill. These things do not go into the product, but they determine whether the agent can avoid detours.
In the comments, someone added a more practical perspective: setting a fixed time to look back over this week's conversation logs is more composed than emergency firefighting. Someone else told the truth: I have always wanted to do this, but when Friday comes I just want to lie down. This friction is real; going through dozens of session logs is not easy work.
Other people do it differently: at the end of each session, have the tool proactively ask whether to save the current memory, then periodically compact and clean up. Spread the organizing across each conversation to avoid a mountain piling up by Friday. Two rhythms, each with its own costs.
One unsolved problem: how to automatically get Codex conversations. So far, apart from official export and manually using a plugin, there is no smooth way. Automation in this area is still unsatisfactory; anyone who wants to do a weekly review has to build a workflow themselves first.
The whole thing sounds like working for AI, but from another angle, it is more like turning one-off improvisations into accumulable assets. Conversations are one-time; skills and notes are not. The difference is whether someone is willing to sit down on Friday night and fish them out of the chat window.
发布时间: 2026-11-08 23:39