How to change one thing on a design in an AI image tool without it quietly redesigning the whole piece. This came up on the bench again this week, so here is the full method. Keep one reference image as the truth. Pick the single image you are working from and stay on it. Every request you make is a change against that one image, never against a description of it in words and never against a second reference image. The moment a second image enters the conversation, the tool starts blending two aesthetics rather than adjusting one. Name the change as one axis with a number. "Make the panel smaller" is not an instruction, it is an invitation to redesign. The tool has no idea what "smaller" means to you, so it picks. "Panel length down by one third, width unchanged" is an instruction. It names the one thing that moves, the one thing that does not, and by how much. Same piece, one change. Ratios travel as numbers, never as a second picture. If you want a proportion carried across from something else, do not show the tool a second image and hope it understands you mean just the ratio. It will not. It will pull in whatever else is sitting in that second image, usually the ornament and the general look, not only the proportion you wanted. Write the ratio down as a number instead. Generating and editing are two different jobs. If you keep asking for a small change and it keeps coming back as a bigger redesign than you wanted, stop regenerating from the prompt and switch to an edit operation on the actual pixels of the image you already have. An edit modifies what is there. A regeneration makes something new that merely resembles it, and small requests get lost in that gap. Check the thing you did not touch. When you use a reference purely for a proportion, watch that the tool has not quietly dragged the reference's own ornament or detailing back into your piece along with it. Look at the parts you never asked to change, not just the part you did. None of this needs a fancier tool. It is about how you talk to the one you already have.

This week's three. Comments are not applause. I spent time this week studying a brand positioning coach on Instagram. 1,862 followers, 66 posts. Her median post earns 30 comments against 37.5 likes. The industry average is roughly 0.02 to 0.05 comments per like. She achieves 0.81 by ending every caption with "comment [keyword] and I will send you my [resource]". Comments are how the post gets delivered to people, and they are how she captures every lead. Her best-reaching post (949 likes) was her worst-converting (22 comments). She treats reach and revenue as separate instruments, and builds for the one that pays. AI video briefing: remove the ambiguous state. I rebuilt a video clip three times this week. It showed a rolled scroll of paper unrolling. First attempt: paper paid out of both ends of the roll. Second: the roll resized itself mid-motion. Both caused by the same thing: I left the model to decide the starting geometry. The fix was starting the clip mid-unroll. No opening state, no wrong choice. The instruction that finally held named one thing moving, one thing still, and named the impossible moves so they could not happen. Stop photographing the product. An audit of fine-jewellery Instagram accounts this week produced one line I have not stopped thinking about. "Stop photographing the product, start photographing the world it belongs in." Three named visual environments set up for the brand: a Persian miniature interior, a calligraphy ground, a textile. The piece enters the world. The world is what people return to see. This is the last one of these before Wednesday. The free session is 19 August, seven in the evening UK time, online, at no cost, nothing to sign up for beyond turning up. I will build on a real attendee's business live and take questions, not a pitch with a demo bolted on. If you are coming Wednesday, what is the one thing about your own business you would want built on, live, if it came up?

I have been building rather than posting. That stops today: a post here every Sunday, one thing from the lab each week. Three things from this week. Two delivery promises on the same page. I went through the Silux site properly and found that 33 of its 115 product pages stated the lead time twice, and disagreed with themselves. The spec table said four to six weeks. The product's own description, further down the same page, said something else, up to ten to twelve weeks in one case. Both visible. Same page. Been like that for months and I never caught it, because I never read my own page the way a stranger does. Fixed now: one figure per page. If you run a shop, that is worth checking on your own site. It rarely gets caught, because nobody reads the whole page in one go except a stranger deciding whether to buy. Getting an AI render to change one thing. When you ask an image tool to tweak a design, it usually redesigns the whole piece instead. What works: keep one reference image as the single source of truth, and describe the change as one named measurement with a number attached. "Length down by a third, width unchanged" gets you the same piece with one thing altered. "Make it smaller" gets you a new design that is vaguely smaller. The first three seconds. Finished my first proper explainer film this week, six minutes forty-eight. The opening took longer than everything else combined. On YouTube the first three seconds decide whether anyone stays, and text on a black screen is not a hook. Coming up: the free session is Wednesday 19 August, more on that nearer the time. Which of the three above do you want me to go deeper on, or does your own site actually state a lead time anywhere at all?