Veo UGC Workflows, Codex Migration, 1Password .env — AI Daily May 02
780 messages · 70 active members
Overview
Topics
Kieran detailed his production stack: Veo for talking UGC, Kling for silent b-roll, 8-second scenes chained via ffmpeg with a consistent reference image, submagic for captions — skipping last-frame workflows due to high failure rates. Seedance 2 wins on realism for ecomm trust signals but distracts from proven Veo ad volume. Mandarin prompts bypass English content filters, and kie.ai is the best non-technical onramp now that Google Flow's UI is 'absolute horseshit'.
@jasonakatiff hit Codex rate limits fast (15% of weekly xhigh in 2-3 hours) with 45+ min job times vs 10-15 min on Claude Code. Codex needs explicit instructions and should be treated as a tool, not a peer; symlink .claude→.codex and CLAUDE.md→AGENTS.md is the standard bridge. The hidden /goal beta lets you fire-and-forget with a success definition, and Grok 4.3 jumped to #2 on AA Intelligence Index (53) but with an 8-point dip in non-hallucination rate.
Jason migrated hundreds of vars across 15+ files to 1Password's new environments feature (public beta), which replaces .env.local with a 1Password-backed mount. Sav310 pushed back recommending Infisical for machine-first/agent workflows, successfully mapping 262 sprawled keys into Docker and Cloudflare via Agent Vault. Frequent auth prompts when Claude accesses secrets remain a UX pain.
@jasonakatiff runs 12 pen-test libs and is budgeting $15k each for real pen testing and SOC2 audits. @geilt argued prompt-based audits cover only basics and that AI tends to bolt on unnecessary layers like JWT over already-robust RBAC — infra architecture matters more than code. Separately, @jcartu advocated an uncensored Qwen3.6-35B (NVFP4) on a 5090 via vLLM hitting 160tps as a cloud-Claude-Code replacement.
A new free skill (bradautomates/claude-video) splits any YouTube/Loom/Instagram video into frames + Whisper/Groq transcripts so Claude can analyze it for ~$1 per 45min. @tounano and @samtome converged on running most 'E2E' tests in happy-dom with React mounted in the test runner — reserving real browser E2E for 2-3 happy paths and drag-and-drop. @jasonakatiff also shared a Notion writeup on the agent improvement loop using OpenTelemetry traces, scoring, and offline evals.
Key Takeaways
- For UGC video ads, chain 8-second Veo scenes with ffmpeg using a consistent reference image and skip last-frame workflows; Mandarin prompts bypass Seedance's English content filters.
- Codex needs different prompting than Claude Code — explicit step-by-step instructions, treat it as a tool not a peer; symlink .claude→.codex and CLAUDE.md→AGENTS.md.
- 1Password's .env environments beta is production-ready for local dev; use Infisical for machine/agent workflows.
- Real SOC2 + third-party pen testing runs ~$15k each, and AI-generated code often adds unnecessary security layers — 'works' ≠ 'needed' or 'secure at scale'.
- Run most E2E tests in happy-dom with React mounted in the test runner (95%+ speed gains); reserve Chromium for 2-3 happy paths and drag-and-drop.
Hot Threads
Video model recommendations and Veo workflow — Kie.ai, 8-second scene chaining, Seedance vs Veo for UGC
1Password .env migration, Infisical, AWS vs GCP vs Vercel for prod secrets
AI security, SOC2 compliance, and whether all human-written software needs rewriting