Session 18: Nano Banana + Veo 3.1 Multi-Shot Video — UGC Genie, an AI-UGC Product Built Live

Pre-AI UGC: brands found creators (often via Fiverr), shipped product, and paid for review-style videos posted on both accounts. AI disrupts every input at once: video/voice models are good enough that audiences accept AI creators 'if it adds value — if it's sloppy, they're not interested'; one operator can run a whole roster of AI-creator accounts; and the brand-side pains — not established, no influencer connections, no clout, contract breaches, conflicts of interest — all disappear when the 'creator' is a workflow. 'With AI, it's a level playing field... if you crack the workflow, it becomes repeatable, and you can post as much content as you want.'

The commercial proof is named: icon.com (New York) — '6 human-grade UGC ads, $1,000 a month... it's literally our workflow. We're honestly just replicating exactly what they're doing.' The session's product IS the market analysis.

Three layers, drawn on the whiteboard: PROVIDERS — the model owners (Nano Banana by Google, GPT Image by OpenAI, Kling, Seedance by ByteDance), each behind its own account and API. AGGREGATORS/CONSOLIDATORS — Higgsfield: one platform, one MCP/API over many image/video models, freestyle enough to make 'cinematic episodes' (the K-drama series example); fal.ai named as the interchangeable alternative, and the modularity doctrine applies ('if you don't want Nano Banana, change it'). CLONING APPS — HeyGen: purpose-built for AI avatars and voice cloning (now with UGC features), the answer when the character must be a specific real person — which is exactly Paul's client-replica question, answered 'it's possible... now it's a free hand.'

The engineering consequence: build against the aggregator boundary and keep it REPLACEABLE (the README's own design principle) — 'in the backend you can have fal.ai, Higgsfield, or Veo itself... whichever is cheapest executes, and you don't have to tell your end user.'

The product shape: upload zone → creative-direction prompt (one sentence: 'a creator discovers this product, demonstrates the most useful benefit, gives an honest, energetic recommendation') → server-side workflow engine → node canvas with per-node lifecycle (pending/queued/running/completed/failed) → generated video in an in-app player. Higgsfield MCP runs server-side (the agent itself flags that a deployed web app can't reach a local MCP — 'the MCP bridge must run server-side'). The agent invents a CREDIT-FREE PREVIEW MODE purely because the prompt said learners were watching — transparency requests become features.

The live run is the proof: Syed's coffee photo → real Higgsfield generation (credits visibly drop 1487→1412; 75 credits, 5-15 minutes) → the Kenco ad, narrative fully AI-written from the one-line prompt. Scope is held by VERSION DISCIPLINE: V1 shipped, V2 = 'I want to see a real output' (upload reliability, functional magic-wand via OpenAI, audio on/off switch that flows into Higgsfield's generate-audio setting), V3 = the parking lot (YouTube auto-post, multi-preview comparison, per-platform settings, character selection). Honesty stays on: sidebar pages are dummies, reload loses state ('it doesn't have a proper backend just yet — add Supabase'), and the Vercel deploy is labeled a public demo because the credits belong to Outskill.

The loop: open the dashboard in Codex's IN-APP BROWSER ('this is ChatGPT Atlas — they integrated their browser inside ChatGPT itself'), hit ANNOTATE, click the exact component, type the complaint ('this feature is great, but it's not fully working'), send all annotations as one batch — each lands in the chat bound to its element, and the agent's mental model summarizes the batch as 'seven refinements as one cohesive interaction pass.' Attribution is cheerfully conceded: 'Lovable did it first, but Codex copied it.' The payoff sentence: 'developers used to sit down — this used to be a task that took 8 hours to build and ship. Now it's just a prompt.'

Supporting cast: the /side chat for status without interrupting; the agent's own visible cursor doing a live browser check ('that's not me — I'm hands off'); and the diagnostic honesty the loop surfaces — the agent finds its own upload bug ('the entire drop zone is implemented as a label containing other interactive controls — that breaks browser patterns'). Token discipline rides along: the agent's own 'take the metro, not the rocket,' and Dileep's caution repeated — loop engineering gets you A to B 'but it might just use a rocket ship.'

The pricing doctrine, delivered as Q&A: DON'T sell an API layer — 'a lot of people do API and it just becomes undercutting each other; the margins become very low.' DO run it as an agency: retainer ($1-2k/month) for platform access with usage limits — 'it's basically SaaS, but not software as a service: SERVICE AS A SOFTWARE.' NEVER freemium for generation products: 'anytime they get free video generation they will rush to your platform... freemium users are never sticky.' And the endgame he keeps circling: as models commoditize ('you can build your own n8n at this point'), the durable asset is the INTERFACE plus the HARNESS — sell access as a license ('$100 per year, nothing extra — you're selling the product, not the codebase') and let customers BRING THEIR OWN KEYS, because 'a lot of people are gonna use their own Claude, their own ChatGPT as the main form of interaction.'

The open-source posture is the same bet from the other side: the $10k codebase ships MIT precisely because the code was never the product — the workflow knowledge, the interface, and the service around it are.