Let Grok Bot Run Its Own Performance Review; Author Says Usage Fell 30%

Avid posted a prompt on X called DOGE, short for Department of Grok Efficiency. It does not teach a bot how to do the work; instead, it has a bot periodically audit itself or other bots: which tasks overlap, where things get stuck, what is dead weight, and then write a report. According to the author, this prompt cut his Grok Bot usage by 30%; paste it tonight and by tomorrow morning every bot can fix its own mistakes.

The screenshot shows a Markdown file whose title is simply DOGE. The file first defines an identity: an efficiency auditor for AI bots, responsible for finding issues that need to be deleted, merged, or fixed and reporting them to the user. The first conversation asks only a few things: time zone, audit frequency, prohibited projects, and whether to stay quiet. Each subsequent audit covers an inventory list, change log, overlapping tasks, bottlenecks, dead weight, and cost estimates. Restarts, sync and recovery operations, normal daily usage and planned work, and blacklisted bots all count as noise and are not reported.

Reports are required to be clear and concise. After the user answers, it saves useful files and instructions into the bot it takes over and confirms changes after making them. It sends a brief summary once a week, including the total number of bots, total number of routines, and daily run count. Every audit must be logged so runs can be compared and duplicate work avoided. The file ends with a Never list: changing settings without permission, sending messages on the user's behalf, and fabricating data.

The author claims this prompt will find what each bot is doing wrong, rewrite the weak parts of its instructions, and prove on real tasks that the fix works. One commenter replied that making the bot prove fixes with real tasks is the only part that will not be skipped. Another commenter warned that you should save the old instructions first, because sometimes the rewrite makes things worse and you need to roll back. That warning is more worth looking at first than the 30%.

The 30% is Avid's own claim; the tweet does not give a statistical basis, so treat it as a number to be verified. DOGE is not magic; it writes an audit process and permission boundaries into the prompt. What really determines the result is the audit frequency, the blacklist, whether changing settings is allowed, and whether a human reads the reports. The safe way to use it is to try it on one bot first, back up the old instructions, set the audit frequency, initially let it only read reports, and only allow it to change settings after confirming the fixes work.

In the same quoted tweet, Avid also listed a bot pipeline: Scout scans Product Hunt, GitHub Trending, HN, Reddit, and X, and writes trending.md; Pain Miner pulls public comments and complaints, scores sentiment, and writes a ranking of pain points into pain.md, using only public data and not making up quotes. Then there are Competitor Mapper, Behavior Reverser, Strategy Lead, CEO, Spec Writer, Builder, and Video Director. The prompt is not isolated; behind it is a division of labor. In a reply, the author said they used this system to release Serein, a free Mac browser that gives Claude Code, Codex, Cursor, and Grok real tabs that can be viewed, taken over, and stopped. It comes from the same source as DOGE, but it is not the core of this article.

If you have a bunch of bots running, what is worth trying is the audit and rollback mechanism, not the 30% number. Save the old instructions first, then let DOGE speak.

Original link: https://x.com/Av1dlive/status/2116527254390067489

Avid's article: https://x.com/i/article/2111086043319701505

Serein: https://serein.pro