Jordan built a chat-based AI image editor in a day to test continuity, then we built a sales deck in Gamma in about five minutes. What worked and what didn't.
AI can now edit a photo of you and keep your face, but only mostly, and the results degrade as edits stack up. On Episode 10 of Built This Week, Jordan Metzner showed a chat-based image editor he built in a day that put him in a brown cowboy hat on a ranch from a quick screenshot, and Sam Nadler walked us through building a three-card sales deck in Gamma in about five minutes. We also covered Google's Nano Banana model, Anthropic's browser agent and what AI deck tools mean for consultants.
Jordan says everybody is trying to replace Photoshop with AI, and it hasn't happened yet. Photoshop and the rest of the Adobe suite carry twenty to thirty years of software development, and professionals still use them every day. Even Figma, for all its success, hasn't replaced them.
Part of the reason, he says, is continuity: you ask AI to change an image, and the final image should have the changes but still be the same person.
So if I ask something like, put a cowboy hat on Sam, I get a cowboy hat on Sam, but I still get Sam. I don't get another face that looks kinda like Sam.
— Jordan Metzner, Episode 10
Jordan says continuity has been a challenge for most AI models, and a few recent releases, including Google's new Nano Banana model, have made it easier to describe the change you want and get it.
Watch Jordan's AI image editor demo in Episode 10 (starts at 3:58) · Episode 10 notes and transcript
Jordan built a small chatbot-based image editing tool for the episode. He sees it as useful for small businesses, especially ecommerce, that constantly need product shots and headshots. Sam pointed out that Jordan built it between one day and the next, and said the point is to move fast and give our teams access to useful tools quickly, even if not everything is perfect right away.
The demo started with a screenshot of Jordan in an office-style virtual background. Sam, who has been thinking about starting a cowboy hat brand from Nashville, asked for a brown cowboy hat on a ranch. Using the Kontext Pro edit tool, the app returned Jordan in a brown hat on a ranch. Sam noticed the lighting: the hat cast a shadow across Jordan's shoulder that looked fairly natural, though the image still looked a little AI generated.
Each new request takes the latest image as the root. They added a western button-down, then hot pink accents and a bolo tie, then turned Jordan into a cartoon character. Jordan warned that this is "an image of an image," so he expects some continuity degradation over time.
For a business example, Jordan asked the tool to put a green glass beer bottle in his hand. It worked. Asking it to show him drinking the beer was harder, and the first try didn't land. He called the whole thing a little prototype showing what these models can do, and noted that OpenAI has had image generation for a while, but not continuity.
He also raised the risk. In the wrong hands, he says, this could be used dangerously: what if the person in the photo doesn't drink alcohol? The real question becomes whether an image is real.
So, maybe this looks artificial today, but, it may not look artificial tomorrow with just, one or two more improvements.
— Jordan Metzner, Episode 10
Sam admitted the live demo was a bit of a flop and asked what the hype was about. Jordan described it as an upgrade to Google Gemini's image capabilities. Google's examples show it merging two images, such as a person and a dog placed together on a basketball court, and interior design uses like adding a couch or a wall to a room.
Jordan noted the model in Gemini may differ from the one he was running through his own application layer. His bigger point is that this is now at the level where developers can build it into apps. He expects to see it on iPhones and Android phones and inside ecommerce apps that help people picture things before they buy them.
He also said Google's Veo 3 is still the best video model he has seen, ahead of OpenAI's Sora, and that video is even more exciting than images to him because of how important it is for social ad creation.
Anthropic is getting into browser agents, following Perplexity's Comet and OpenAI's agent. Neither of us had tried it yet; Jordan noted it was in limited release. Sam said his default search is now usually an LLM, so having one directly in the browser makes a lot of sense to him.
Jordan's example was booking an international flight. He normally keeps several tabs open across times, days and airports, then checks his points. He imagines writing some rules about how he likes to travel and letting a Claude agent open ten or fifteen tabs on Google Flights or Expedia to find the best deal. Today's chat experience, he says, mostly searches and hands you a link out, and it is not clear how comprehensive that search was. Sam had used AI that same day to shortlist hotels about an hour outside Seattle, and said the results were pretty close to what he needed.
We also touched on Will Smith's tour video, which viewers say used a fake AI crowd with deformed faces. Jordan called it the beginning of artists using AI as trickery. Sam expects public figures to use AI to influence their audiences to some degree.
Sam described Gamma as a tool that uses AI to speed up building presentations and websites. He had used it only a couple of times over a month and a half of subscription. For the demo, we asked for a three-card sales deck about Built This Week aimed at potential sponsors.
Gamma first produced an editable outline of cards with titles and bullets. From there you pick the amount of text per card (Sam suggests concise), a theme, an AI image model and a style; we chose line art. You can then click an image and ask AI to change it, which we used to make one more realistic. Jordan was surprised that Gamma surfaced information about our podcast from the internet without being told anything about us. It can present directly or export to PDF, PowerPoint or Google Slides, and Sam thinks it has about 20 templates, including a pitch deck.
Sam mostly uses it for professional one pagers, without needing a graphic designer.
You can definitely get a skeleton going within seconds that is probably 65, 75, 80% of the way there.
— Sam Nadler, Episode 10
He says work that may have taken hours drops to maybe an hour and a half. Jordan, who worked at Amazon and its writing-heavy culture, doesn't make many slide decks, but called it a great tool for pitch decks and sales decks. He is surprised Google, OpenAI and Anthropic haven't done a better job with AI in slides, and sees Gamma as the market leader so far.
Sam asked what LLM research plus tools like Gamma mean for management consultants. Jordan said he is no expert in that space, but the time consultants spend making decks is going to zero. He hopes the research side of the job stays important, while noting ChatGPT is an amazing researcher.
He shared an anecdote from a trial attorney he spoke with over the weekend. As the attorney told it, his litigation firm is hiring fewer data researchers, mostly because tools like ChatGPT can now do research on cases, and that has changed how the firm thinks about future hires. Jordan sees the same kind of shift reaching education, schools and VC pitch decks.
See every edit in the cowboy hat demo and the full Gamma build on camera in Episode 10 of Built This Week, where the full transcript is also on the page. The next week, in Episode 11, we took the same idea to music with an AI DJ mixer.
