One Solution to Your Dwindling Usage Limits

If you have been struggling with usage limits, I've found my own solution, and 1000 other people on here have done something similar. So add this to the list. Put plainly, Sol is overkill for most work. If you are building a relatively simple app, website, etc., you aren't solving "complex" problems. You are solving problems with known solutions, and just don't know the appropriate code to get the job done. That's Terra or Luna work. Sure, Sol will delegate, but not before blowing a bunch of tokens and filling the context window only to figure out that it's not the right guy for the job. Start with the fundamentals. Sit down with ChatGPT and have it help you write meaningful Agents.md, Soul.md, User.md, Memory.md or whatever other scoped documents you need. Make the agents you use delegate tasks appropriately, understand your preferences, and don't let Sol waste time/tokens reading irrelevant files and pondering the meaning of life. Since taking some time to do that, it has (seemingly) solved my usage limitation issues. Hermes has been a great tool for implementing this kind of stuff too. Example : I'm currently refactoring a custom AI harness for legal workflows. It got a bit convoluted because I got lazy and just said "Go" for several turns like an asshole. I thought I had a detailed and robust plan in place and that was enough—lesson learned. Before, I'd set it to Sol-Extra High and hope for the best while watching my usage limit drop 5-10% per turn. Yesterday, I spent a few hours getting my docs back in order, with ChatGPT mostly asking me questions and making suggestions along the way—tedious but helpful. Now, almost every workflow uses Sol as the planner-orchestrator-architect-reviewer while delegating most implementation work to Luna and Terra subagents. I do have some MoA evaluation flows set up too, but those are case-specific uses with 3x3 agent limits. I also directed the agents to minimize comments where the code's purpose is obvious or where they are just adding unnecessary context. I'm 2.5 hours deep into some fairly complex rebuilding and I've dropped 2% since I started. Initial testing showed massive improvements to the end product, more consistent structure, and even better memory efficiency—working with building the harness for local models, so efficiency matters a lot. I can't recommend doing the same highly enough. Your wallet will thank you. TL/DR : Update your docs with ChatGPT and dictate appropriate agent usage. You'll go far and save money Style Note : The em-dashes were placed in this post intentionally by a human being because they know how to write well. May we all embrace such newfound enthusiasm for punctuation.