Grateful to @Nils Davis for starting a separate thread on that - I want to create a consolidated version here based on some research, best practice and trial and error on how to optimize your Claude usage so you don't run out of tokens. So if you've been hitting your usage limits, here are some tips to try out and to keep in mind. Quick context: a token is roughly one word. Every time you send a message, Claude re-reads your entire conversation from the top. So message 1 is cheap. For message 30 Claude is re-reading 29 previous exchanges before it even looks at your new question. That's why your credits disappear. 1. Convert files to markdown before you upload them. A single PDF page costs 1,500 to 3,000 tokens. Screenshots can be 1,300 tokens each. If you upload the same 15-page PDF to 4 different chats, that's 180,000+ tokens gone on one document. Open a Google Doc, paste the text you need, download as .md. Markdown is the love language of LLMs. 2. Use Sonnet for everyday work. Save Opus for the heavy stuff. Opus burns tokens 5x faster than Sonnet. Grammar checks, brainstorming, reformatting, short answers. Sonnet handles all of it at a fraction of the cost. If the task takes Claude less than 30 seconds to answer, it doesn't need Opus. Switch before you start. It takes 2 clicks. 3. Turn off extended thinking when you don't need it. Extended thinking burns through your allowance way faster than you'd expect. It's extra steps, extra outputs, extra compute. If you're not working on something genuinely complex, turn it off. 4. New task = new chat. Every message in a thread carries the full conversation history. A 20-message session burns roughly 105,000 tokens. A 30-message session? 232,000. If you went from writing a LinkedIn post to drafting a client proposal in the same chat, Claude is still re-reading the LinkedIn stuff every time it thinks about your proposal. 5. Be specific from your very first message. "Summarize this document" followed by "actually just the financial risks in section 3" is two expensive messages when it could've been one. Tell Claude exactly where to look and what to do. The more specific you are, the fewer tokens you burn.

How many of you have a solid Brand DNA docs that you use when working with LLMs? By Brand DNA I mean a rich context document (or a folder with docs) about your business - things like your founder story, ICP, offers, long-term strategy, client feedback, tone of voice, brand values etc. Basically a living document you keep enriching over time and feed to AI tools so they actually get you right off the bat. This is going to be more and more important with Claude Cowork and I wonder if everyone got it covered or if I should create some resources for how to create one?

Ok so I know that many of us relate to being multi passionate business owners ( or aspiring!) and I just want to share this podcast episode that I found so deeply resonating and helpful in accepting this as a superpower. I agree with Dan on so many levels and really admire his ability to synthesize so many diverse aspects and insights. Highly recommend to listen/watch when you have time! https://youtu.be/ExNWGF-q64M?si=LLZ5di5MYUbmM_Az

Hello everyone! I’ve been building a simple tool to help solopreneurs become more aware of their wellbeing by looking into how their day actually felt. The latest version of the Wellbeing Radar GPT (WRG) prototype is ready, and I’d love a few volunteers to try it for one week. What it does In under 90 seconds, WRG gives you a quick daily check-in across: - Hope - Energy - Focus - Manageability - Exhaustion And offers one small optional “next move.” How to use it 1. Open the WRG GPT 2. Click or Type “Start” 3. Complete the check-in daily 4. Type “Show Results” What I’d love to learn After the week: - Was WRG useful in your day? - Did it help you notice anything meaningful? - Did it meet expectations? - What could make it better? Even a few lines of feedback would be incredibly helpful. Thank you so much for supporting this project! It truly means a lot. Karl

Hello everyone! I hope your week is starting well. A little while ago, I mentioned that I was experimenting with a prompt-engineering technique called Verbalized Sampling. Here’s a quick, simple explanation. Verbalized Sampling (VS) is a training-free method that pushes AI to reveal several possible answers with probability scores, instead of giving one safe, generic reply. By asking the model to “show its distribution,” you unlock ideas it normally hides—especially the more creative, low-probability ones. 👉 Why this matters: Most prompts drive the AI toward the most predictable answer (mode collapse). VS lets you see the full range—from standard to unconventional—so you can choose better, sharper, or more original insights. 🌟Advanced Version of VS I’ve also included an advanced version called Tail Sampling, which focuses specifically on the rare answers (the “probability < X%” zone) for those who want more contrarian or breakthrough ideas. I uploaded a short guide with examples and use cases. If anything is unclear or you’d like more demonstrations, just let me know! I’m happy to dig deeper. If you try this, let me know how it went. Have fun with it!