Let code check before the bot thinks

@Solomonrojie posted a Grok Bot setup yesterday that boils down to one line I like: "files remember. code filters. Grok Bot finishes."

The idea is simple. Keep the working chat short. Save research and decisions in files. Let scripts do the boring checks, like whether a source changed, before the model spends any thinking on it. Fresh research only starts when nothing saved answers the question, when an input changed, or when a fact has expired.

His post also quotes token savings for this setup. I haven't measured those on my own runs, so I'm not repeating them here.

Why this matched what I already run

I posted on October 19 that I find myself using Grok Bot for everything and hadn't opened a terminal in a month. I used to keep five to ten terminals open at all times. Once the bot is doing that much, the waste shows up fast: the same page fetched again, the same failing call tried again, a chat that gets so long every new turn drags the whole history along.

My setup handles that with shared skills that every agent can load. One says to check cheaply whether a feed, inbox, repo or URL changed before fetching or reasoning about it again. Another caps retries, backs off, and never repeats an identical failing call. A third keeps long chats small when a task closes, and another filters big tool output down to the lines the task needs before it enters context.

That's the same order Solomon describes. The files and the code do the remembering and the filtering. The model only gets the part that needs judgment.

The part I'd add

His tree says to stop a failed operation after two attempts and keep the work that already finished. My retry skill follows the same rule, and it's the piece of his setup I'd copy first.

Credit to @Solomonrojie for the post.