My Claude Code workflow no one knows about
TL;DR
The real workflow is idea → context → design → analytics → experiments — Amir shows how he connects Idea Browser, Claude Code, Paper, Tailark, and Humblytics to go from an AI sales-coaching concept to a live landing page with A/B tests in roughly 30 minutes.
Context is the moat, not just the prompt — his key point is that most builders can vibe-code a page, but very few preserve evolving business context like ICP, positioning, growth strategy, and experiment history so agents can make better decisions over time.
Good AI design comes from taste and references, not magic prompts — instead of saying “improve the design,” Amir feeds Claude screenshots, style guides, and Tailark components, then uses precise constraints like “subtle animation” to get polished results instead of generic purple landing pages.
Custom-coded sites plus agents are replacing no-code CMS workflows for fast growth teams — Amir says Humblytics migrated from Webflow to custom code so Claude can act as the CMS, create pages, update content, run tests, and connect directly to Google Ads, Meta, Stripe, and analytics tools.
The terminal is becoming the interface for work — Greg and Amir revisit their old “Cursor is the future of work” thesis and update it: now the terminal, with MCP connections and long context windows, is where software, design, analytics, and growth operations are actually getting done.
There’s a live arbitrage opportunity here for marketers and agencies — Greg compares this moment to early Facebook ads at 5-cent CPCs, arguing that almost nobody knows this stack exists yet, which creates room to charge $5k-$20k/month for personalized landing pages, CRO, and campaign ops.
The Breakdown
The promise: no more ugly vibe-coded landing pages
Greg opens by pushing Amir to “give all the sauce,” and Amir sets the frame clearly: this episode is about taking an idea, validating it, refining the design, building a polished landing page, and then using data to optimize for revenue. The subtext is obvious and funny: they’re aiming to avoid the all-too-familiar “purple vibe-coded landing page” outcome.
Idea Browser as the business memory layer
Amir starts with Idea Browser’s new Claude Code MCP integration and argues the missing piece isn’t idea generation — it’s tracking how a business evolves over time. Using an example they built earlier with Jordan — an “AI sparring partner for B2B sales teams” focused on freight software — he shows how files like offer definition, talk tracks, ICP, and growth strategy give Claude the context to make smarter next-step decisions.
Building the lead magnet from stored context
From that project context, Amir uses an Idea Browser skill called “lead magnet legend” to generate a niche offer: “Five objections that kill freight software deals.” He likes that it saves the output back into the project files, because that turns one-off prompting into a running company brain you can revisit later.
Why Paper matters between Figma and raw code
The next handoff is into Paper, which Amir describes as the missing middle layer between static Figma mockups and coding directly in Claude. His pitch is simple: if everyone is building in code now, Paper gives designers and marketers a place to iterate visually, create variations, and keep track of what changed without losing themselves in the terminal.
How he gets polished design instead of generic AI output
Amir’s actual design trick is refreshingly practical: he gives Claude screenshots of sites he likes, asks it to extract the design language, and turns that into a reusable style guide. He insists vibe-coded design isn’t inherently bad — his own polished Humblytics site was built this way — but only after refining in Paper and borrowing taste from libraries like Tailark, which he uses for clean prebuilt sections and subtle UI patterns.
Tailark, subtle animation, and the art of prompting taste
He installs a Tailark component into the lead magnet page and uses it as a design reference, then explains a detail that feels very true in practice: wording matters. Instead of broad prompts like “improve the design,” he tells Claude to make “subtle” refinements and “subtle animation,” because agents respond much better to guardrails than vague creative freedom. There’s even a random ant on the table while this is happening, and they decide the ant can stay and “get the direct sauce.”
The terminal as the new operating system for growth
Midway through, the conversation zooms out into a bigger thesis: the terminal is becoming the interface for work. Greg and Amir connect this to MCPs, agent wallets, markdown pages for crawlers, and Gartner’s prediction that 20% of internet commerce by 2030 will be done by agents; Amir even says more agents will visit websites than humans.
From live page to live A/B test without deploying code
Once the lead magnet page is pushed live, Amir brings in Humblytics to track scroll depth, bounce rate, click maps, and form submissions, then launches a no-code A/B test on the homepage headline directly through Claude. The wild part is that the variant goes live without shipping new code, and the broader pitch lands hard: if you connect ads, analytics, experiments, and site generation into one stack, you can run personalized landing pages and CRO like a tiny team with an army behind it.