✨ The AI Quirk: You give AI a clear instruction at the start of a prompt… but the response seems to ignore it completely. Even stranger, if you repeat the instruction later in the prompt, suddenly the AI follows it perfectly. ✨ What’s Going On: - Large language models weigh instructions "based on proximity and clarity" within the prompt. - Instructions buried early in a long message can lose influence once the model begins predicting the response. - The model often prioritizes "the most recent instruction signals" it sees. - If a prompt contains mixed signals (examples, context, and instructions together), the model may treat the first instruction as "background instead of a rule". Example: You start with: 1) Write this in bullet points. 2) Then provide a long paragraph of context. The model may treat the context as the main task and default to paragraphs. But if you end the prompt with: “Use bullet points for the final answer”, the output suddenly follows the rule. ✨ What To Do If You See It: - Place "critical instructions at the end of the prompt". - Separate instructions from context using spacing or labels. - Repeat important constraints when precision matters. Try this prompt: “Using the context above, produce the final answer in bullet points only.” ✨ Why This Happens: AI isn’t reading instructions like a human would. It’s predicting the next most likely text — and "AI tends to pay the most attention to the instructions it sees last." ✨ AI Bits & Pieces — helping people and businesses adopt AI with confidence.

I’ve been cruising along lately, just loving life with Claude Code, but I recently got absolutely humbled by Cowork. I’ve actually made up a new acronym for the event: WTT. The reality is, I’m not quite productive enough on Code yet to justify dropping $100 or $200 a month in tokens. So, I’ve been managing my token "allowance" like a kid who gets $5 a week when candy bars cost $2. Everything was going great until a buddy asked me to show him the ropes on Cowork. Being confident (and perhaps a bit too proud), I said, "Sure!" I didn't mention I’d never actually run this specific workflow before. I figured—how hard can it be? I set up a scheduled task to roll through my Gmail account and give me a daily briefing. Awesome. Easy. Total "Pro" move. So, I’m showing off, doing this very impressive demo for him, and I hit "Run Now." I’m leaning back, like: "Look, you can see it thinking... it’s reading the email, producing the brief, preparing the summary..." DONK. 💥 OUT OF TOKENS! WTT (What the Token?!) And then it hits me: I cannot Code. I cannot Chat. I cannot Cowork. #WTT Thinking on my feet, I started to explain, "You see, managing tokens is just a sophisticated part of the LLM lifestyle," all while seething under the covers. So, that was yesterday. I decided to shrug it off, tough it out, and wait for the rolling reset while I slept. ☀️ The next morning, I was all fired up to get back into Claude Code to build my slick new app. I sat down, opened the terminal, and... DONK. 💥 OUT OF TOKENS! #WTT Then It hit me... 🤖 Claude Cowork had woken up before I did. It chewed through my entire brand-new allowance of tokens just before I was all fired up to "Claude Code.". I cannot Code. I cannot Chat. I cannot Cowork. #WTT The ultimate irony? I couldn’t even grammar-check this post in Claude. I had to go over and do it in Gemini! LOL. #WTT @Matthew Sutherland @Nick Mohler @Usman Mohammed @John Romano

Welcome to the Out of the Box series — where I explore what can be built with no-code and low-code AI tools in 30 minutes or less. No manuals. No tutorials. Just curiosity and creation in motion. This time I revisited Sora 2 a few months later to see how the experience has evolved. App: Sora by OpenAI Time: Under 30 Minutes Category: AI Video Creation / Prompt-Directed Video Video Title: Move Over Rover, The Dog Days of Coding Are Over - Claude Code is The Cats Meow 🎥 What Is Sora? Sora is an AI video generation platform that transforms a simple text prompt into lifelike, cinematic scenes — complete with motion, lighting, and visual storytelling. Think of it as having a director, camera crew, and editor… all powered by a prompt. ⚙️ Experience 1 — The First Test A few months ago, I ran an Out of the Box experiment with Sora using a simple presenter-style scene. The results were impressive for early generative video, but the workflow still felt a bit like experimentation. The outputs were interesting, but not something that added much practical value beyond demonstrating what the technology could do. If you’re curious about that original test, you can see the full post here: 👉 https://www.skool.com/ai-bits-and-pieces/out-of-the-box-in-30-sora-2?p=e63f6633 That first experiment helped show what was possible, but the bigger question was how quickly the experience would evolve. ⚙️ Experience 2 — Revisiting It Today For the second experiment, I tried something completely different — a playful, high-motion scene designed to test character behavior and storytelling. Prompt theme: A cat driving a quad runner at high speed — Fast & Furious style — with a dog riding on the back howling and clearly terrified. The twist: - The cat is labeled “Claude Code.” - The dog is labeled “ChatGPT.” Experiment 2 Video: https://sora.chatgpt.com/p/s_69b4d4703dbc819180c914a61747c81f?psh=HXVzZXItQWI5dFRpa3JRS1RTSmhwbDY3VlFYaWxv.4nGp4ZY9Gsxo