Codex Desktop Wins, UGC Video Harness, Opus 4.8 Slow — AI Daily Jun 05
662 messages · 76 active members
Overview
Topics
Multiple builders reported Codex Desktop is god-tier for browser automation — logging into webapps, clicking confirmation links, and verifying frontend changes using a real Chrome profile instead of detectable Playwright. One user compressed 3 weeks of CLI work into 1 week. Claude still holds an edge on frontend design, but sentiment is shifting fast and Codex's speed has eliminated the old 'wait and brain-rot' downtime.
@expadz detailed a fully automated UGC pipeline using Claude Code Desktop: ffmpeg frame extraction (1fps), ElevenLabs word-level timestamps for ~2.5 words/sec pacing, and Claude vision auto-retries clips 3-5 times until QA passes. Pairs with a small ~50-clip pre-approved B-roll library (not 1000+) that Claude picks from, with ffmpeg background removal and auto subtitles. Veo3 generates ~200 clips/hour; Omni isn't worth the extra cost for talking-head UGC.
@mb29266 shared his shotgun playbook: one CBO with adsets by avatar/concept, ~1000 bidcapped ads per batch (200 copies × 4-5 images), kill anything that doesn't spend in 5 days. Images first because volume is cheap, then ~75% of winning images get converted into videos. 95% of ads get zero impressions but cost is now trivial.
Opus 4.8 ran painfully slow today even on medium effort, and a mid-morning Claude API outage threw 529 Overloaded errors (incident fprlnsvdnr2k). @tounano keeps Opus 4.5 as his interactive-planning sweet spot paired with Codex for implementation; Codex Spark is fast but too dumb. @GuruTime noted you can swap 4.8's thinking model to speed it up. Codex xhigh + ultraplan is the new hard-task default.
@geilt showcased a Redis/Valkey message-bus harness solving multi-provider routing pain (especially Kimi) and a fully automated 8GB Mac mini media pipeline with remote Ollama qwen3:8b classification, ClamAV scanning, and Plex hardlinks. @Coybh flagged Elixir scoring ~98% on LLM code-gen benchmarks, with its bus architecture fitting agentic design. Meanwhile builders are deferring Mac Mini/Studio buys ahead of Apple's June 8-12 event hoping for M5 silicon — 64GB+ Minis have 6-month lead times. @sav310 got an LLM Council running via Discord bridge with a synthesizer agent.
Key Takeaways
- Codex Desktop is the new default for browser automation — real Chrome profile bypasses Playwright detection and one builder compressed 3 weeks of CLI work into 1 week
- Build a verifiability harness for UGC video: ffmpeg frame extraction + ElevenLabs word-level timestamps + Claude vision auto-retry, paired with a small ~50-clip B-roll library (not 1000+)
- Test ads with images first — they're cheap to generate at volume, and ~75% of winning images convert into winning videos, so don't burn Veo3 compute on unproven concepts
- Opus 4.8 is too slow for interactive planning today; Opus 4.5 remains the sweet spot, and Codex xhigh + ultraplan is the new go-to for hard tasks
- Skip MCPs for Meta Ads and similar APIs — modern models handle raw API docs faster and more reliably than slow MCP wrappers
Hot Threads
Verifiability harness + pre-approved B-roll library for automated UGC video ads
Codex Desktop vs Claude Code CLI for browser automation and full-stack verification
Bidcapped CBO structure for launching 1000+ ads per batch