UGC Video Pipelines, Remote Claude Code, Cloudflare Pages — AI Daily May 11
600 messages · 96 active members
Overview
Topics
Kieran and @rutty123 shared full agentic pipelines: script + accent + image ref → GPT 12-panel storyboards → Seedance 2.0/Kling/Veo with nano banana references → auto-QA for wrong words and dupes → merge + Submagic captions, all running through Claude Code. Realism wins came from reference image quality, skin blemishes, and iPhone-style framing — not model choice. Pronunciation bugs in Seedance persist; prompting in Chinese surfaced as a workaround for content restrictions.
@tounano, @arielletolome and @iannagy debated the levelsio Hetzner/Claude Code approach vs local M4 development. Consensus: remote makes sense for long-running agents and heavy Docker/RUST builds, while Beelink N100 home servers work for steady locations. @arielletolome shared his full TMUX vibe-coding Notion setup.
Multiple builders confirmed migrating from Vercel to Cloudflare Pages for landers, PBN sites, and B2B traffic — citing the $5 plan's range, built-in D1/R2, and EmDash CMS integration. Vercel still preferred for full apps with Supabase form capture. Tip: give your agent a Cloudflare API token for end-to-end deploys.
@mb29266 shared his agent pipeline picking from a massive template library to one-shot lander copy, planning multiple Claude Max accounts running 24/7. @c_1media noted Claude easily pushes aggressive DR copy while GPT-5.5 'acts like it's going to hand-serve you papers from the FTC.' Codex 5.3 still praised as the 'American dev' counterpart for structured engineering work.
Parakeet TDT 0.6B v3 on Together AI flagged as the fastest STT on the artificialanalysis.ai leaderboard. Groq's Whisper Large v3 Turbo and Salad's $0.10/hr transcription API compared as alternatives, with reminders to run your own accuracy tests rather than trust third-party benchmarks.
Key Takeaways
- For one-shot UGC ads: script + accent + image ref → GPT storyboards → Seedance/Veo/Kling with nano banana references → auto-QA loop → Submagic captions. Realism comes from reference quality and iPhone-style framing, not model choice.
- Run long-lived AI agents on remote servers (Hetzner/Contabo) — local dev gets crushed by RUST compiles and Docker builds, and 24/7 agents need uptime you can't get on a laptop.
- Cloudflare Pages is the new default for landers and PBN sites; pair with EmDash CMS and give your agent a CF API token to deploy end-to-end. Keep Vercel for full apps with Supabase.
- Use Claude for aggressive direct response copy and Codex/GPT-5.5 for structured engineering — they behave like European vs American devs, with GPT-5.5 being far more risk-averse on marketing copy.
- For Seedance content restrictions, prompt in Chinese (via Mitte's YouTube workaround). For speech-to-text speed, Parakeet TDT 0.6B on Together AI currently leads.
Hot Threads
Automating GPT storyboard → Kie/Seedance/Veo video generation pipeline
levelsio Hetzner/Claude Code remote dev setup vs home servers and M4 local
Best 2026 hosting setup for cold-traffic landers (Vercel vs Cloudflare)