Founder Stack Ep03 with Shai Krishna: Why Wrappers Make AI Slop - CLAUDE.md as the Brain and DESIGN.md as the Art Direction, Founder-Led UGC with AI Avatars, Trial Reels, and Email Journeys

Shai's own first product, Designfolio (a portfolio builder for designers), was hand-coded with ChatGPT as an assistant, not an agent (an assistant helps; an agent takes the initiative to finish the task). To replicate a product he now pastes its URL into a temporary chat (no memory, no bias) and asks for the tech stack and the four features that make an MVP - for Designfolio: add a project, choose a template, publish. The stack answer is Next.js and TypeScript with Tailwind and shadcn on the front end, TipTap for the rich text editor and Supabase behind it; he stresses that founders and marketers can ask the model what each name means. Supabase is the backend he recommends for any MVP: the free tier holds 50,000 monthly active users and 25 USD a month stretches to about 100,000, and only after that does a move to AWS-style infrastructure make sense. His own reverse-engineering point: if you want to copy something in the market, copy the link, ask your favourite model, and use your free model for the thinking.

The live comparison: ChatGPT writes a prompt for a Designfolio-like builder, Shai pastes it into Replit (output: website, no backend, 'build only the builder') and in parallel into Claude Code in a new VS Code window with the Claude Code extension enabled. Replit's result is acceptable but misses most of the flow and one screen only; it has no persistent brain - the chat is the only memory, and a CLAUDE.md given to it is read inconsistently. His explanation for the miss, repeated in the Q&A: a wrapper has to spend the model's credits running itself, so the full potential of the AI is split between coding and wrapper duties; give all agentic credits to the product instead. The same MD file can be handed to Replit, Bolt or Lovable but they skip lines (he has seen it hundreds of times). Claude Code can be used three ways - the desktop Claude app, the VS Code plugin (his favourite: code on the left, so you grow comfortable seeing code without needing to write syntax) and the terminal. Setup in the room: Claude Pro (20-24 USD), VS Code, Extensions, search Claude Code, trust the folder. Opus at high effort for the base build; /usage shows consumption and reset times. He adds that Claude Cowork is now inside the Claude app, that Codex versus Claude is a toss-up (Astra, GPT 6.1 and Opus 5.5 all good) and that he was unimpressed by Gemini 4 for Three.js.

The recipe: in a temporary ChatGPT chat, paste the target URL, ask for the tech stack and MVP features, then ask it to write a CLAUDE.md for that product (Shai at first had a prompt, then replaced it with the file). Create CLAUDE.md in the project's root - Claude Code reads it on every prompt - paste the content, save, then instruct 'read CLAUDE.md and build it', optionally 'no back end for now, just the user flow'. The MD format works for any LLM; agents cannot see, they read text. A habit he recommends: write CLAUDE.md in phases (for his game, phase 5d is invites and phase 5g is player connection requests); when a phase is complete it moves to an archive file so the live CLAUDE.md stays short, and the file doubles as a Trello/Linear-style tracker so you always know what is next. He also warns about a stray CLAUDE.md on the Desktop being picked up in the live demo. The cohort checks that you need not use ChatGPT: you could use Codex or any model; his split is GPT (free) to think, Claude to build, and DeepSeek or Gemini to think if you pay for both coders.

First the failure: a Spider-Man image plus a one-line prompt produced a flat, ugly game (and Claude pulled a memory he did not want). Then the method: (1) in ChatGPT describe the game ('someone roams a 3D city, nothing else, Three.js') and get a CLAUDE.md; (2) paste Valorant and Overwatch screenshots and ask for a design.md - creative direction, palette, bold art direction, what it is not (no generic procedural environment, no realistic city simulator); (3) in Claude Code ask it to align the CLAUDE.md and DESIGN.md to the Valorant feel and, importantly, to import DESIGN.md into CLAUDE.md (the confirmed line: the art direction lives in DESIGN.md and is loaded in every session) because Claude reads CLAUDE.md and ignores other MD files; (4) build with Opus at high effort. Sonnet is enough for the step that only edits context. The same rule for a SaaS product: write 'like Linear, like Pinterest, like LinkedIn' into DESIGN.md. The finished game he shows (from the same two files) has lighting, shadows and a day/night selector. The three differences he lists: Replit gives one screen; CLAUDE.md gives the whole flow; DESIGN.md gives the flow without the slop.

The problem he solves: founders with full-time jobs cannot sit in front of a camera daily and will burn out. HeyGen (heard 'Hagen' / 'agent') builds a video avatar from a short recording (he says 15 seconds in one place and a lit, steady 50 seconds in the Q&A; face and hands look slightly AI), and its memory improves with use (his 7th-8th video in pro mode beat his first). Script: paste the product link into ChatGPT and ask for a reel script (he notes this is not how to write real hooks - use a hooks skill). Pricing as stated: 600 credits a month, about 3,000 rupees or 32-33 USD for roughly 18 videos. Voice: HeyGen's own voice is ~50 percent like you; ElevenLabs gets 70-80 percent and improves as you generate and rate it; no model reaches 100 percent. Editing: drop the avatar video into a folder, open the Claude app on that folder and tell it to use his personal skill ('tasteful HQ' - a DESIGN.md for video: isometric visuals, Apple-style motion, scenes locked). Rules in the prompt: keep the source aspect ratio, add no greens or colour changes, use your own computer's hardware to render (otherwise it renders on its server and takes over two hours), and show sample screenshots before rendering. The run produces a kit of timed animation edits (for example at 0:30 insert this animation) and one command renders the video, about 12-14 minutes of work. He shares the skill zip on Drive; to install it, Claude app, Customize, add, upload - double-clicking it does nothing, and it can also be installed in Codex or Claude Code. He says Remotion and Higgsfield credits were not needed for this path.

His account structure: one account for the product journey (he talks only about Claude because he sells a course on it, so people associate him with Claude), plus a top-of-funnel account built on shareable avatar content ('Claude for designers') that reaches strangers; you mention the product only occasionally and convert followers later. Build in public: the pre-launch post about submitting Tap Away to the App Store reached 490 people before launch day and 30,000 views by the end; look at Insights (views), not likes, because familiarity makes the next post easier to try. Trial reels (Instagram): a reel posted as a trial is shown only to non-followers, so you can test a message outside your niche audience (his profile reaches designers; Tap Away is for everyone) and publish to the main feed only what works; with an avatar tool you can make seven variants with a different shirt, hook, model or avatar. His Designfolio funnel numbers as stated: 60+ new logins a day, 50-60 paid conversions every three weeks, 3-4 percent conversion, about 20 percent retention helped by an email journey, 31,000 active and 41,000 signed-up users. Why founder-led: hired influencers bring 200 logins but nobody remembers them; domain authority and a visible human team do; he advises doubling down on avatars and your personal journey; a follower is not a customer, so keep them in your community (he uses ManyChat for resource automation). Face-swaps only with AI-generated humans, never real people.

His production stack for Designfolio: Next.js, Dodo Payments (heard 'DoorLoop' / 'Duo') for the payment gateway, and Featurebase (heard 'feature based') to collect user questions and feedback. Dodo is a merchant of record that uses Stripe or PayPal underneath and also has an MCP that connects to Claude Code ('integrate the payment gateway, create a product'); his advice is to use it where Stripe is not available and Stripe where it is. Email journeys: the product should email at every state change - waitlist join, welcome, publish ('your portfolio is live'), purchase receipt, subscription expiry, and in his game social messages. Resend (developer email, 3,000 emails a month free) has an if/else-style automation builder; AutoSend (Indian company, by the Peerlist founders, paid only) calls them automations; Resend calls them campaigns; Brevo's free tier (9,000 a month, from the chat) and self-hosted open-source mailers on your own server are alternatives. To integrate, tell Claude Code you want Resend and paste the API key; bot detection can swallow welcome emails sent to temp addresses.

For Tap Away (a replica of the US 'Brick' - an NFC tag that blocks distracting apps) he has no photography skills, so product images come from GPT Image. For 3D he gave photos to Claude and asked it to create a GLB (a 3D model file; GPT image alone gives only a picture, and Codex can also generate them). The GLB then feeds the landing page's in-browser animation (the device rotates as you turn your phone, using device-orientation) or an animation model: the rotation shot came from Higgsfield, the tap shot from Kling (3-4 free videos; Higgsfield has no free tier but is cheap), Meshy was named as another route. The landing page was assembled with Claude Code, a Higgsfield MCP and GPT Image: 'you can use these three to build non-slop designs'. Lighting and time-of-day shadows on the page also came from CLAUDE.md and DESIGN.md.