maxstudio: a decision manual an agent can runmaxstudio：写给 agent 执行的个人决策手册

Nine Markdown files, four Claude Code slash commands, one 38-line shell script, and no application code anywhere. That is the whole repo. It exists because I have roughly ten spare hours a week and kept spending them on whatever felt most interesting that evening. The fix I tried was to write the judgment down first, then hand it to an agent that reads it back the same way every time.

The file that took the most work is a data baseline. I pulled the creator-dashboard numbers for all 36 posts on my Xiaohongshu account and worked out what one new follower actually costs: 567,000 impressions, 83 followers. Two posts both counted as hits — one turned 18,600 impressions into 20 followers, the other turned 264,000 into a single one. So the content rules come down to four formats I'm allowed to write and a list of things I'm not. Thumbnail click-through is where the numbers hurt: my eight highest-CTR posts gained four followers between them, and my single biggest gainer sat below the account median. That reversal is one of two commits whose only job was to overturn a rule the repo had written less than a day earlier — the original text called 17–31% CTR "the hit range," and it now reads as a pass mark, not a contest.

The project the repo currently tracks is a measurement question: when is multi-agent collaboration worth it? The rig runs on pi-feishu-im, an open-source extension I wrote that drops several coding agents into one Feishu group. How a run is set up, how batches are launched, and which files an agent may touch are all fixed in writing. The first controlled comparison came back against my hypothesis: on the same task, the agents working together were slower than one agent alone, and the output was no better.

The routines the repo defines have mostly never run. The reviews folder is empty. The 7:38am topic scan is registered with launchd and does fire, but the script it calls was never made executable — the fix has been sitting in the handoff to-do list since the day it was written — so every run since August 12 has failed, both log files are zero bytes, and the topic inbox is blank. The one finished draft is still marked unpublished. The measurement project's latest conclusions, 163 lines across two files, have been sitting uncommitted since August 14.

The decision layer is real and I use it. The automation layer is a shell that was never plugged in. Rule six of the repo called this in advance: maintaining the system is not progress. Only three things count — code landed, content published, a real person talked to.