Turning "generate one image" into a creative pipeline you can hand off — real client workflows pulled apart on screen
Hacking AI for creative operations at scale ft. Rory Flynn | Config 2026 · Jenny Shi
21 min total·Actually worth watching closely: ~11 min·3 must-watch clips
- 0:12 – 3:20Listen
Creative inflation: why doubling output makes it worse
Opens on where the industry actually is — channels and variants are a bottomless pit, AI lets everyone do more, and demand grows faster still. That sets the terms: point solutions don't fix this.
The line isn't whether you can make one good image. It's whether you can do it at scale — and that only comes from systems.
Mostly argument; the slide is a title card carrying the words. Fine to listen to while doing something else▶ Jump to 0:12Speaker · Rory Flynn - 3:20 – 5:00Watch
Individual output isn't the bottleneck
An extreme example: two images plus an idea and a couple of common tools, and in 30 minutes you have a full typeface — upper and lower case, ligatures, special characters, vector files. Twenty dollars now buys assets that are indistinguishable from reality.
Generation is cheap. What's scarce is the structure to organize it.
The finished typeface is the entire proof here — described out loud it sounds like an exaggeration; seeing the character set and the files on screen is what lands it▶ Jump to 3:20Speaker · Rory Flynn - 5:00 – 7:13Listen
It's only three things: what you give it, how you brief it, how you iterate
Takes the mystique out of "AI creative system": you don't need to know every tool. There's context, what you provide it; direction, how you brief it; and iteration, what you do once something comes back.
All three hold across tools and across media — the tools change, these don't. So learn this first instead of chasing tools.
Pure spoken argument; a three-word frame you'll remember on one pass, and the screen adds nothing▶ Jump to 5:00Speaker · Rory Flynn - 7:13 – 9:40Watch
Deconstruction: breaking a photo into a formula you can fill in
The first core skill. Reverse-engineer a photo down to its non-negotiables — lighting and the rest — because controlling those means controlling the image, and image editing reduces to keep this, change that. It doubles as a demo of what a correct workflow looks like: build the body of the car, apply the design, take multiple angles, then use that as a reference image for every real-life variation.
Don't jump from the starting image to the final ones. Work in steps, piece by piece, with quality control at every step — the way you already do.
The car sequence evolves step by step; watching what each stage adds to the last is far clearer than hearing it described▶ Jump to 7:13Speaker · Rory Flynn - 9:40 – 12:10Listen
System prompts: writing a brief for the model
The second core skill. Like any creative brief: who it's acting as, what it's going to receive, what it does with what it receives, how it should output, and what not to do. Written once, it holds every time, and it passes information from one piece of the tool to the next.
The point isn't getting AI to write your prompts — it's clicking once and using the tool over and over with a different image and a different input.
This is about the structure of the prompt; the screen is mostly walls of text, so listening beats reading▶ Jump to 9:40Speaker · Rory Flynn - 12:10 – 14:15Skim
Workflow development, and building something you give away
The third core skill: which tools you need to get from start to finish, and in what order. Once it's standardized and packaged, you change the input and the whole system runs itself.
Built scalably and structured appropriately, 95% of the time these don't have to change — so you hand it off to someone else and go solve the next problem.
One glance at the node diagram is enough to see how things connect; the narration carries more than the picture▶ Jump to 12:10Speaker · Rory Flynn - 14:15 – 16:40Watch
From one to 400, and BarkBox replacing the shoot
Batch takes only two changes: tell the system prompt to write four prompts instead of one, then add a text iterator that splits them into individual inputs so all four generate at once. Then the BarkBox case — the monthly subscription box shot, moved from studio to digital compositing.
Four could be 400 — and pick the small, frequent problem right in front of you instead of trying to solve way too big a problem.
The compositing flow is assembled on screen one node at a time; the order of placing, flattening and adding lighting only stays straight if you watch it. How the finished image then fans out into banners, UGC and social is also carried by the visuals▶ Jump to 14:15Speaker · Rory Flynn - 16:40 – 21:09Watch
A pitch in four clicks, and how the system spreads
The toy licensing case: four system prompts run the whole thing. 2D vector to 3D, into the batch generator for nine angles with seam logic and texture mapping so it stays consistent, then a scale ref, then back through the batch generator for the studio shots. Then SharkNinja — localization across hundreds of products and 35 markets, after which the systems spread to other lines of business.
Things come out unproportional because there's no context for how big something is. Feed it a scale ref and that reference image becomes the new feeder image for everything downstream.
Every link in the four-step chain has its output on screen, and the before-and-after around the scale ref is the one to look at — that comparison is what makes the step read as necessary rather than redundant▶ Jump to 16:40Speaker · Rory Flynn