💡Let me start with the truth of things first, because the truth is the lesson. ================= Token Warning ⚠️========== Before anyone comes for me 😅 📝 Note: depending on what you are loading, you are accruing token cost. And token cost can ballon if you are not careful. If any of the files in the initializing phase or handoff phase are large this can become a token intensive practice so when curating this process, BE INTENTIONAL only load what is absolutely needed for the project to start up and for you to get to the fun part, building! I budget for my own needs, for me this initialization for me is between 5k and 10k, I have made the mistake of making 40k, don't do that! If you want to understand an estimate of what your character to token ratio is use the site below: Tokenizer - OpenAI API Anthropic tokenizer has not been released yet, but Its a few % higher than OpenAI in some cases. Plain English sentence (115 characters) - OpenAI GPT-4: 23 tokens - OpenAI GPT-4o: 23 tokens - Claude (content only): 23 tokens - Result: identical Code and markdown snippet (86 characters) - OpenAI GPT-4: 24 tokens - OpenAI GPT-4o: 26 tokens - Claude (content only): 27 tokens - Result: Claude a few percent higher Numbers and punctuation (55 characters) - OpenAI GPT-4: 25 tokens - OpenAI GPT-4o: 25 tokens - Claude (content only): 27 tokens - Result: Claude about 8 percent higher One honest footnote: Claude's API adds roughly 7 fixed tokens per message for formatting. On a single short line that looks like a big gap. On a real document it amortizes down to almost nothing. ✅The takeaway: for the same English text, the tokenizers land within a few percent of each other. The scary "Claude costs way more tokens" headlines do not hold up when you actually measure it. Verified beats confident, every time. You have been learned and you have been warned! ============ Back to it! 👇=================

Every time you drop a PDF or Word doc into a model, you pay tokens for formatting it never reasons over. Layout, styling, table scaffolding. None of it earns its place in the context window. So I built to-md. It turns .pdf, .docx, .pptx and .xlsx into Markdown, and writes two files: a faithful one, and a lean one stripped to the minimal text. The lean file is the one you paste in. One engine, three ways in: a command line tool, a Finder right-click action (Adds the path of the lean file to your clipboard too), and a Claude Code skill that fires the moment you hand it a document. The principle underneath it: the model should only read what it has to think about. The rest is waste you are paying for. Pre-process the document. Then prompt. //A<3