@joshuablatch: Claude reads my whole video now for about 3,000 tokens, because my Mac does everything else for free. Read frame by frame, 40 seconds of video is roughly 50,000 tokens, and I was basically paying @claude to stare at my face. Now my Mac does the looking first with what's already built in: text recognition, face detection and a cutout that works like a green screen. Whisper writes the transcript and ffmpeg turns the whole video into one contact sheet, all running on the Mac. By the time Claude opens anything, it's reading a one page summary and at most two images. The most important piece is the audit. The prompt starts by having Claude go through your own sessions and list where you spend the most, so it works even if you never touch video. Here's the prompt. Follow and drop a comment if you have any questions. Build me a token-saver mod for Claude Code. Start by auditing my setup for token efficiency: 1. Look through my recent sessions and list the 5 tasks where you spend the most tokens. For each, tell me what you were reading (images, video frames, long files, transcripts, web pages) and roughly how many tokens it costs each time. 2. For each one, tell me whether my Mac could do the heavy lifting locally, for free, before you ever read it. Check what's built in first: Apple's Vision framework (text recognition, face detection, person cutouts) through a small Swift script, Spotlight (mdfind) for finding files, and sips for image details. Then free local tools: ffmpeg for one contact sheet instead of separate frames, Whisper for transcripts, and simple scripts that summarize long files. 3. For the top two, build a local prep step that saves a one-page summary and at most two images, so you read that instead of the raw files. Show me a before-and-after token estimate for each, and wait for my OK before installing anything. #claude #claudecode #ai #tokens

Josh Blatchford

Open In TikTok:

Region: US

Friday 06 November 2026 00:02:26 GMT