I spent my last job in venture capital reading pitch decks for due diligence, and the thing nobody tells founders is this: the reviewer is a human being before they're a rational one. Before I judged the product, before I judged the team, I'd already decided whether I wanted to keep reading. A stale deck gets closed before the strategy is ever read. So I rebuilt a real client deck — one that went on to close a high-ticket deal — and made the whole process into a video. Here's the short version, so this post is useful even if you never click. THE 3 TELLS 1. The vertical accent bar on every card. border-left: 3px solid. It's the single most recognizable thing AI makes, and once you see it you can't unsee it. 2. The grid of rounded boxes. I counted 11 on one slide. Every slide being the same card grid makes a deck read like a form — and nobody argues with a form, they fill it in and leave. 3. No texture at all. Flat white or flat dark, nothing else. THE 4 BLOCKS THAT FIX IT This is the entire prompt structure. Works in Claude Code, Codex, Gemini — it's a way of thinking, not a tool. 1. Guardrails — the don'ts and dos. Start here, not with what you want. You're not describing a target, you're fencing off the average. AI's default output IS the average of everything it's seen, so most of the work is closing doors. 2. Experience — what should the reader feel, and who is the reader? 3. Design — the concrete layer. Colour, theme, template, layout, font. 4. Assets — links, filepaths, screenshots, images. Block 4 is the one people skip and it's the one that matters most. I wasted an embarrassing amount of time typing "warm linen texture, analog, handcrafted" and getting garbage back every time. AI has never touched linen. It has no idea what that word means as a feeling. You can't adjective your way there — you show it the photo. The gradient in this deck came from a phone photo I took at a startup pitch event in Seoul. Badly, at an angle, with the back of someone's head in it. That photo was worth more than every adjective I could type.