Nov 26, 2025 · 59m · startup-ideas
Reviewing Claude Opus 4.5
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Greg Isenberg and James (The Boring Marketer) run a live, head-to-head coding benchmark comparing Anthropic's Claude Opus 4.5 against Google's Gemini 3 Pro by building a full-stack SaaS application from scratch. Through hands-on landing page generation, interactive prototyping, and an analysis of custom Claude skills, they explore the evolving frontier of non-technical software development.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 22.9% of the talking time here. How this is scored →
Greg as informed peer 4.5 Guest teaching 3.9 Guest disagreement 0.1 Greg pushing back 0.7
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 10:16 James corrects host assumption on active skills
James gently corrects Greg's assumption that custom copywriting skills generated the copy, noting only the design skill was active.
Hardest push from Greg ▶ 25:59 Host delivers direct negative critique of ad designWhen asked for an honest evaluation of James's ad, Greg directly refuses to flatter it, pointing out that it is too busy and has too much text.
Biggest teaching moment ▶ 46:00 Systematic breakdown of Claude skill developmentJames provides a comprehensive masterclass on configuring custom Claude skills using Perplexity MCP research on industry leaders combined with brand voice distillation.
Greg holds their own ▶ 55:20 Host frames content as the foundational UX elementGreg demonstrates product expertise by reframing why vibe-coded apps fail to convert, explaining that words are the actual UX.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Episode Overview and Experiment Setup | 4 | 2 | 0 | 0 | Greg sets the experiment parameters by selecting a startup concept (EstateClear) from IdeaBrowser, while James sets up the side-by-side prompt testing between Claude Opus 4.5 and Gemini 3 Pro. The dynamic is fully cooperative and exploratory. | |
| Sponsor: IdeaBrowser.com Promotion | 4 | 2 | 0 | 0 | Following Greg's IdeaBrowser sponsor read, James explains why he runs the test in Google AI Studio to isolate variables, and both make friendly predictions about model behaviors. | |
| Reviewing Claude Opus 4.5's Landing Page Output | 3 | 3 | 1 | 0 | Greg asks whether custom copy skills generated the landing page text, and James clarifies that only the front-end design skill was used, explaining that Opus produced the copy natively. | |
| Comparing Gemini 3 Pro's Landing Page Build | 5 | 2 | 0 | 1 | Greg points out that Gemini included an AI copywriting widget just as he had predicted, evaluating its layout and visual aesthetics against Opus 4.5. | |
| Prompting Full Clickable Prototypes and Backend Architecture | 3 | 5 | 0 | 0 | Greg inquires about building full SaaS backends, prompting James to explain his production stack using Neon, Clerk, Vercel, and Stripe, comparing Claude's end-to-end execution to other models. | |
| Analyzing Google's Vertically Integrated Ecosystem | 5 | 4 | 0 | 0 | James reviews Google's vertical integration across hardware TPUs, workspace data, and developer tools, while Greg adds business context regarding Google's equity stake in Anthropic. | |
| Exploring Nano Banana Pro and AI Ad Generation | 6 | 3 | 0 | 5 | James shows his Meta ad creatives generated with Nano Banana Pro and asks for honest feedback. Greg delivers candid pushback, critiquing the ad as overly busy and text-heavy before highlighting its workable visual elements. | |
| Evaluating Claude Opus 4.5's Interactive App Prototype | 6 | 4 | 1 | 1 | James questions the recurring use case of an estate probate tool, and Greg elaborates on vertical SaaS mechanics, retention hurdles, and customer discovery strategies. | |
| Reviewing Gemini Prototype and Debugging Strategies | 3 | 6 | 0 | 0 | When Gemini generates console errors, James educates Greg on his prompt debugging framework, including using terminal 'ultra think' commands and prompting specialized sub-agents to trace root causes. | |
| Testing Google Anti-gravity IDE and Browser Integration | 5 | 5 | 0 | 2 | Greg wagers that Google's Anti-gravity IDE will yield a worse design output, while James explains the value of Anti-gravity's Chrome extension for direct DOM inspection during debugging. | |
| Anti-gravity Output Review and Workflow Insights | 6 | 2 | 0 | 0 | The test validates Greg's design prediction, but both note Anti-gravity's autonomous use of Nano Banana to create mockups first, extracting UI wireframe lessons. | |
| Mastering Claude Skills and Elevated Direct Response | 3 | 7 | 0 | 0 | James delivers an in-depth walkthrough of building custom Claude skills for 'elevated direct response,' combining automated research via Perplexity MCP with personal writing voice distillation. | |
| The Claude Front-End Design Skill and Copy Optimization | 6 | 5 | 0 | 0 | James demonstrates Anthropic's official front-end design plugin, and Greg synthesizes the discussion by asserting that content is UX and copy drives vibe coding conversion rates. |
Statements from this episode (20)
Isenberg: A family probate dashboard would have extremely high customer retention
“And the retention is probably. Insane. You're not going to cancel that.”
James uses Perplexity's Model Context Protocol for foundational competitive research
“Typically what I would do is I would do some foundational research, you know, like I would put this into there. Hey, use the perplexity MCP. Tell me who is in this space. What's a unique Kind of differentiating angle that we can take”
Isenberg: Claude will be solid and expected, Gemini buggier and exploratory
“I think that the, my guess is Opus is going to come up, come and do like a very solid, fully featured SaaS app, but it's going to be like more, Kind of similar to what you're saying. It's going to be similar to what you'd expect, but I think it's going to be r…”
Isenberg: Opus 4.5 and Gemini 3 reach new vibe coding stratosphere
“Anecdotally, it does feel like we've, with Opus, 4.5 and Gemini three, it does feel like we've entered a new stratosphere of vibe coding.”
James: AI tools merge code and design into full-stack vibe building
“I think that the lines are just blurring across the board. Like, you know, the line between a non-technical person being able to code blurred, and I would always feel very frustrated with, like, the design portion of this whole process, you know. Yeah, you can…”
Isenberg: Claude 4.5's landing page design objectively beats Gemini 3 Pro
“I think you're right there.
Like, I think objectively it was nicer.”
James: Built production-ready SaaS from scratch using Claude 4.5 Sonnet
“Like with Claude 4.5 sonnet, like I've taken you know idea for a SAS from zero to production ready.”
James: Getting from zero to ready software is easier with Claude
“What I have personally found, and to be clear, I haven't tried to build a full on application with Gemini three yet, but I think that Claude's earlier models, like I would always go back to them. Like when I was trying to use codex and GPT five before 5.1 came…”
James: Google AI Studio's vertical integration provides a massive developer advantage
“Google has the full range of capabilities there that they own that You know, they can integrate directly into like AI studio which I think from just a convenience perspective for most people is a really powerful value proposition. And from a cost perspective, …”
Isenberg: Nano Banana Pro delivers significantly better text rendering
“It's significantly better than the first version. The number one thing I noticed is just text is better.”
James: Creators can find alpha by copying DTC e-commerce ad formats
“They're testing way more ads than any other industry. They're learning really fast, borrowing some ideas from there, and then translating it into, like, you know, sort of the creator world is, I think, there's some alpha there.”
Isenberg: Finding workflow and social dynamic is the hardest part of vertical apps
“The hardest part of building a vertical app is figuring out what is the workflow, figuring out what is the social aspect and then obviously figuring out what is the vertical.”
James: Claude Opus 4.5 will resolve bugs with less sub-agent prompting
“I'm thinking that its ability to,
Work through the bugs without sort of getting lost is going to be a lot stronger.
and I wouldn't maybe need ultra think or sub agents or maybe even those additional agents on the side.”
Isenberg: Claude Opus 4.5 beats Gemini 3 Pro on initial app builds
“But I think if you look at, based on the two prompts that we've done, I think it's clear that estate clear on 4.5 beat Gemini three.”
James: Google Anti-gravity's DOM access will minimize back-and-forth debugging
“So with the Chrome extension that anti-gravity connects to, it can access that data for you programmatically. So it's accessing the DOM. It's seeing sort of the data behind the scenes. So you know, I suspect it will greatly minimize the back and forth in this …”
James: No competitor matches Google's interconnected developer and advertising ecosystem
“I mean, like the ability to stay interconnected and grow lifetime value. No one can match that with this ecosystem that they have, which is, it's wild.”
James: AI Studio Produces Better Single-Prompt Design Output Than Anti-gravity
“I think, you know, my take would be outside of this image, it feels a little basic, so I think your instinct that AI Studio will have a better initial design output from one prompt was correct.”
James: Upfront Claude skills and research yield better results with single prompts
“You can get way better results with less prompting and less work in between. So do the research up front, set up the skills, Grow one prompt into the system, and you'll get a great result. So doing the boring stuff up front, if you will, doing the research, bu…”
Isenberg: Vibe coders fail because they neglect conversion copywriting
“The big piece that people miss with vibe coding is they vibe code stuff, but then the copy sucks. So then people, you know, they're converting at one percent or less than one percent and they're like, why? And then, and they come to the conclusion that, oh, vi…”
The Boring Marketer: Claude Opus 4.5 One-Shots Interfaces With Front-End Skill
“4.5 Opus, I think is going to be an incredible model just based off some of these initial tests that we did and that I was doing last night. When you combine it with that front end design skill, it's able to one shot some pretty incredible interfaces and desig…”