Somebody asked me how to get around the character limit on an agent system prompt. I understand the instinct and I think it is the wrong question. The better I get at this, the more I think about it the other way around. There are agent files, memory files, CLAUDE.md files, and the goal is not to get any of them to capacity. The goal is to keep them as short as possible. The reason is mechanical. Claude Code, Hermes, all of them have system prompts they access during every single turn. So a character you add is not a one time cost. It is a toll you pay on every request for the life of the project. You want those files optimized and short so the thing does its job correctly, and when it does need to go grab a tool or a skill file or a set of rules, it knows exactly where to go with no extra words in the way. So the strategy is not how do I add more characters. It is how do I reduce as many as possible and still tell it to look in the right places for the right things, only as needed. What I actually do: File pruning about once a month. A cleanup pass across the CLAUDE.md files, the memory files and the agent files, shortening all of them. You see a big jump in speed. An end of session routine that asks one question every time. Based on this session, do you recommend any changes to the CLAUDE.md. The goal is to shrink or replace, never to add. Tying the important skills to commands. If the skill fires on a slash command, the file does not need to carry the logic for deciding whether to trigger it, and it is not reading a second document to make that decision. Karpathy has the framing I keep coming back to, and I have handed it to my cleanup skill as a source. The smarter these models get, the shorter those files need to be. Which means you should be changing them all the time. The longer you use this stuff, the more bloated it gets. If your file has only grown since the day you wrote it, it is pruning time.