Most founders spend 3 months building before they find out nobody wants the thing.
Amir, the founder of Humblytics, does the opposite. On Greg Isenberg's podcast he went from a raw idea to a designed landing page with a live A/B test running on it. In one sitting. On camera.
Zero deploys, zero developers waiting on tickets. And the page doesn't have that purple vibe-coded look Greg jokes about in the intro.
And here's the part that got me: companies already pay him $5K to $10K a month to run this exact system for them.
Watch the full episode here, then I'll break down the whole stack:
The stack: four tools and a terminal
Everything in this episode runs through Claude Code in a terminal. The tools plug in as MCPs:
→ IdeaBrowser (Greg's product): holds the idea, the ICP, the positioning, the growth strategy. All as files the agent can read and write.
→ Paper: a design canvas wired to Claude Code. Design to code, code to design, both directions.
→ Tailark: a library of clean UI blocks you install as reference components.
→ Humblytics: analytics, heatmaps, funnels, and no-code A/B experiments.
Amir's line on why this matters stuck with me: "The terminal is the interface of work."
Two years ago people laughed at that take. Now he runs his whole marketing site through it.
From idea to landing page in about 30 minutes
The idea came from a livestream earlier that day with Jordan: an AI sparring partner for B2B sales teams. Reps practice objection handling on simulated prospects and get feedback before they burn real pipeline.
Then they niched it down to freight software sales. Specific beats general, every time.
The build sequence:
→ Pull the project context from IdeaBrowser through the MCP (ICP, offer, competitive positioning, all of it).
→ Run a "lead magnet" skill that turned the positioning into a PDF guide: "Five Objections That Kill Freight Software Deals."
→ Run a landing page architect skill to build the page for that guide.
→ Watch Claude Code design it section by section, live, inside Paper.
Total time for the lead magnet page: about 30 minutes.
Greg's reaction on camera: this lead magnet looks better than 99.99999% of lead magnets that exist.
He's not wrong. And his advice on top of it was the tactical gold of the episode: don't make one lead magnet. Ship a new one every few weeks and keep the winners.
The design trick that kills vibe-coded ugliness
Everyone assumes AI design output looks like AI design output.
Amir's fix is embarrassingly simple: never let the model design from imagination.
→ Screenshot sites you love. Drop them into Claude and say "extrapolate the key design elements and build me a design system."
→ Save that design system as a file. Reference it in every future session so new sections stay consistent.
→ Pull real components from Tailark instead of letting the model invent layouts.
→ Use the word "subtle" in your prompts. Not "improve the design" (too broad). Say "refine the design, keep layouts consistent, keep animations subtle."
That one word, subtle, is doing a lot of work. A vague prompt gives the model room to go weird. The constraint is what makes the result look designed on purpose.
He was honest about the cost too: a properly dialed page still takes a couple hours of refining, not one magic prompt. The demo moved fast because he already had his component library and design system built.
(I went deep on this exact problem in my breakdown of the AI design workflow that doesn't ship slop, which pairs well with this episode.)
A/B tests without touching code
This is where it gets unfair.
Amir copied his Humblytics API key into Claude Code and typed: create an A/B experiment for the headline, suggest the variant.
The agent pulled the page's live analytics first. 40% average scroll depth. 25% bounce rate. Then it wrote a control and a challenger, and launched the test.
The variant headline it came up with: "Every lost deal started with an objection your rep wasn't ready for."
Greg read it out loud on camera and called it "pretty fire." He's right. That's a better headline than most humans write on the first try.
The wild part: no deploy. The Humblytics script swaps content dynamically, so the test went live without pushing code or waiting on anyone. Refresh the page, see the variant, watch the clicks come in.
Most founders never A/B test anything because the setup friction eats them alive. This removes the friction entirely.
The autonomous CRO agent (the $10K/month part)
The live demo was cute. What Amir runs at Humblytics day to day is the real business lesson.
They rebuilt their marketing site from Webflow, to Framer, to fully custom code. Why? So the agent can be the CMS. No page builder in the way.
Then they set up a weekly cron job in Claude Code with three skills:
→ A paid media manager that pulls performance data from Meta and Google Ads.
→ A CRO optimizer that launches and manages A/B tests.
→ A funnel report that connects Stripe and the rest of the stack, then tells them what to fix.
Every ad set on Google gets its own personalized landing page, each with its own experiments. Try doing that through Webflow. You'd need a full-time person just clicking buttons. These business automations are what a growth team of five used to look like, running on a schedule.
And the kicker: customers came to Humblytics and said "we'll pay you $5K, $10K a month, just run this for us."
So the workflow became a managed service. Greg pushed it further: plenty of businesses would pay $15K or $20K a month for this. That's a whole agency model hiding inside one terminal setup.
The 5-cent click window is open again
Greg told a story from the early Facebook ads days that frames why this matters now.
Clicks cost 5 cents back then. He'd go to social gaming companies (FarmVille era) and say "pay me $2 per install." They said yes instantly because they monetized way better than that. Free money, because almost nobody knew the platform yet.
Today that same click costs around $23. The window closed.
His point: this stack is the new 5-cent click. 99.999% of people don't know it exists. If you can generate ideas, ship beautiful pages, and A/B test them in hours instead of weeks, you're arbitraging attention against everyone who still needs a sprint planning meeting to change a headline.
Add the agent wave on top: Gartner projects 20% of internet commerce will be agent-driven by 2030. Amir's thesis is that more agents than humans will visit websites. Whoever ships and tests fastest wins that transition too.
This is the same shift I covered in the rise of the marketing engineer: the person who can do this becomes the most valuable hire (or founder) in the room.
Steal this playbook
Here's the whole system, compressed:
→ Step 1: Write your idea, ICP, and offer into files the agent can read. (IdeaBrowser does this, a folder of markdown works too.)
→ Step 2: Build a design system from screenshots of sites you love. Save it as a file.
→ Step 3: Generate a lead magnet from your positioning, then a landing page for it. Use real UI components as references.
→ Step 4: Wire up analytics and launch a headline A/B test the same day.
→ Step 5: Put the reporting on a weekly cron job so the data comes to you.
→ Step 6: Ship a new lead magnet every few weeks. Kill losers, scale winners.
The founders in my world waste months polishing products nobody asked for. This flips it: validate with a live page and real click data before you write serious code.
FAQ
What tools do I need to copy Amir's workflow?
Four things: Claude Code in your terminal, a design layer connected to it (Amir uses Paper), a UI component library for reference styles (he uses Tailark), and an analytics tool that can run no-code A/B tests (his product, Humblytics). IdeaBrowser holds the business context, but structured markdown files get you most of the way.
Do I need to know how to code?
Less than you think. Amir types plain English prompts and the agent handles the build. Where taste matters is direction: reference designs, constraints like "subtle," and knowing which test to run. Expect a couple hours to properly refine a page, not one prompt.
Why custom code instead of Webflow or Framer?
Because the agent becomes your CMS. Amir's team migrated from Webflow to Framer to custom code specifically so Claude Code could ship changes, spin up personalized landing pages per ad campaign, and run experiments without a page builder slowing it down.
Can this actually make money or is it a demo trick?
It's already a business. Humblytics customers pay $5K to $10K a month to have this run as a managed service, and Greg argues $15K to $20K retainers are realistic. The same system also works as an internal growth engine for your own product, which is the cheaper place to start.
Want the full episodes, not the recap?
I interview bootstrapped founders every week about exactly this kind of playbook: real numbers, real systems, no fluff.
