We shipped 10 cohort funnels in 4 days, mostly with AI

10 funnels. 4 days. Almost entirely AI.

Last week we shipped 10 cohort-class landing-page funnels at /lp/ai-classes/*. Each one is a 22-section sales page with its own custom product mockup, attribution-tracked registration form, three-tier early-bird pricing, ship-or-we-build-it guarantee, and full SEO + JSON-LD scaffolding.

Total time from "let's add a class" to "all 10 funnels live, production build passing": about 4 working days, almost entirely written with Cursor + Claude Code + Claude Agent SDK. Two engineers, sometimes one.

This post is the honest writeup. What worked, what we got wrong, what the 2026 build-with-AI workflow actually looks like when an agency uses it on itself.

The stack we used

• Cursor as the editor. Tab-completion is now better than ours used to be.
• Claude Code for the "give it a goal, walk away for 4 minutes" tasks — usually scaffolding the syllabus copy or generating the product mockup components.
• Claude Agent SDK for one specific job: read the existing 4 cohort classes, generate a new one matching the same shape but for a different topic. Worked once, then we hand-tuned each.
• Next.js 16 App Router + Tailwind v4 + Framer Motion — same stack we ship client work on.

What we built per cohort

Each of the 10 cohort funnels is a discrete sales page with the following sections in order:

1. Hero with product mockup (the iPhone, the browser, the Telegram bot, the Discord channel — each one is custom SVG, no stock images)
2. Three-tier price-ladder strip (early-bird seats 1–10 at ~33% off, mid 11–15 at ~16% off, final 16–20 at headline)
3. Trust strip with anonymized stock-brand logos (per our honesty rules — no fake clients, but Acme/Northwind/Helix-style is industry standard)
4. Instant value — $3,000+ of toolkit, before you pay a cent — the moment that re-frames "I'm holding a seat" into "I'm getting value right now risk-free"
5. Knife-twist agitation, before/after, named instructor card
6. Authority + outcomes
7. Mid-page CTA bar
8. Week-by-week syllabus with deliverables
9. Schedule + stack chips
10. Vs-table (cohort vs YouTube vs hire vs DIY)
11. Price anchor (A Maven cohort is $1,500–$2,500. You pay $497.)
12. Bonus stack with strikethrough $ values + total bundled value
13. Ship-Or-We-Build-It Guarantee — if you do the work and don't ship, we finish your project on a Hayaiti SKU at no charge
14. Triple risk reversal
15. "What happens in the next 60 seconds → 24h → cohort fills → day 35"
16. Founder letter
17. Email-first form (progressive disclosure)
18. Sharp objection FAQ
19. Honest "this isn't for you if…" disqualifier
20. Final CTA
21. Sticky bottom bar that follows scroll

Where AI helped most

• Mock components. The custom phone-frame iPhone with subscription card, the browser frame with the Stripe button, the Telegram message feed with the buy/sell alert — all 10 mockups generated in pair-mode with Cursor in about 90 minutes. Each is React + Tailwind + Framer Motion, no images.
• Sales copy. The "before" and "after" pairs, the agitation, the objection answers — generated as drafts then hand-tuned. Cursor never produced a final draft, but it produced 80%-good drafts in seconds.
• Schema generation. Every JSON-LD blob (Service, WebApplication, ItemList, BlogPosting, Event, JobPosting, etc.) was generated by pointing Claude at the schema.org docs and the page's data shape.

Where AI hurt us

• Pricing research. The first pass priced classes at $995–$1,495. After actually looking at Maven cohort comps + cold-ad CommuniPass benchmarks ($297–$497 is the proven cold-ad-→-cohort window), we re-priced to $397–$697. AI wrote the original wrong numbers because we asked it to mirror Maven's premium tier, which is for warm-audience creator courses, not cold ad funnels. Lesson: AI does what you ask; it doesn't tell you the question is wrong.
• The first version of "Find your SKU". AI happily kept calling the page "Find your SKU" because that's what the URL was. A real human reading the H1 said "which actual company writes find ur sku? this sounds stupid as fuck" — and they were right. We rebranded the H1 to "Which Hayaiti project fits yours?" Lesson: AI mirrors the brand voice it sees in the codebase. If the existing voice is internal jargon, the new voice will be too.

What the cohort program teaches

Every cohort is the same structure: 5 weeks, 1 live build session per week, 1 office hours per week. Cohort cap of 20. Class doesn't start until the 20th seat is reserved. $0 charged at reservation — charged only when the cohort fills. Refunded if it doesn't fill in 30 days.

The 10 topics are a survey of what indie operators are actually trying to ship in 2026:

• iOS app on the App Store
• SaaS website with Stripe
• Crypto trading bot, paper-traded
• AI agent that handles your repetitive task
• Voice AI agent on a real phone number
• Chrome extension on the Web Store
• Discord bot with AI moderation
• AI micro-SaaS hitting MRR
• Fine-tuned local LLM running on Ollama
• AI cold-outbound system (Apollo + Clay + Instantly)

Each is positioned for the audience that actually pays for that outcome: trading bot for the person paying $97/mo for Discord signals, voice agent for the agency owner reselling to local businesses, micro-SaaS for the founder watching Twitter post "$10k MRR in public," and so on.

What's next

First cohort kickoff happens whenever the 20th seat lands in any of the

1. We don't run a half-empty class. The hardest design choice was

saying no to "soft launch with 7 students" — but the cohort experience falls apart below ~12, so we'd rather refund and wait.

If you want to see the funnels, the URLs are noindex (ad-only). But /changelog documents what shipped, and /roadmap documents what's next.