Claude Outage & Design, Codex Mobile, Seedance Ads — AI Daily Jun 23

703 messages · 89 active members

703
messages
89
active members
@mb29266, @Kieran, @arielletolome
top contributors

Overview

Anthropic dominated the day on both sides of the coin: Claude went down mid-morning with 500 errors that broke active CLI sessions, sending builders to Codex, OpenCode, and OhMyPi as alternative harnesses — and later, Anthropic quietly shipped Design directly into the Claude co-work app (bottom-left above account settings). The new Codex mobile app drew strong praise for syncing with terminal sessions via OAuth/SSH so you can edit local repos from your phone, though most builders still prefer terminal for deep work, often pairing Codex with Claude Code and experimenting with GLM 4.5 (reportedly near-Opus quality at a fraction of the cost via OpenRouter). On the agent-building side, @ltv_1 kicked off a long thread on multi-agent orchestration: @sibunting shared a Telegram-based dev/QA stack, @tidemid is pushing toward $10k+/month inference with goal-driven loops holding their own API keys, and @archimorty framed the split between bounded feature work (<2hr) vs. uncertain optimization tasks worth running overnight. @fmill1 shipped a Claude-built ping/post bidder saving ~$3k/mo. The creative/performance crowd reverse-engineered 0xROAS's viral yapper ads (Omni + Seedance, ~$1.40 per 15sec at 480p with a 10/10 first-frame ref and humanized movement prompts), debated a 3-campaign Meta structure (ABO testing → CBO scaling → bidcap graveyard) for an EU supplement brand, and shared tooling for ad spy (Anstrex, Foreplay, VidTao, Tyver) and video QC (Gemini via frame.io, plus SpeechBrain for ~50ms accent detection). Ops threads covered Twilio A2P 10DLC approval, transactional email stacks (Resend, Mailgun, SES, Sendy), and VPS migration off local Mac minis.

Topics

Claude hit 500 errors mid-morning, breaking active CLI sessions inconsistently and pushing builders to status.claude.com. Later in the day, Anthropic quietly shipped Design directly into the Claude co-work app (bottom-left above account settings), alongside continued speculation about Sonnet 5 and the banned Fable model.

The new Codex mobile app syncs with terminal sessions via OAuth/SSH so you can edit local repos from your phone — builders love it for scoped tasks and orchestration on the go. During the Claude outage, OhMyPi (lightweight SOTA) and OpenCode (heavier) emerged as viable harnesses; @jonmacofficial reported zero regrets after weeks fully on Codex, which also has a Claude Code import plugin.

Builders reverse-engineered 0xROAS's viral yapper ads: @swh800 confirmed Omni + Seedance as the stack, and @mb29266 shared the formula — 10/10 first-frame ref image, scripted filler words, prompts enforcing humanized movement, ~$1.40 per 15sec/480p clip. The full course was leaked via MEGA folder.

@ltv_1 asked how builders wire agents for handoffs; @sibunting shared a Telegram-based dev/QA/orchestrator stack with Jira + Obsidian as source of truth. @tidemid is targeting $10k+/month inference via goal-driven loops with API keys and completion conditions, while @archimorty distinguished bounded feature work (<2hr) from optimization/simulation tasks worth running overnight. @fmill1 shipped a Claude-built ping/post bidder saving ~$3k/mo.

Builders debated a 3-campaign Meta structure (ABO testing → CBO scaling → CBO bidcap graveyard) for an EU supplement brand, with warnings that bidcaps can torch budgets during Meta glitches with no refunds. @navuud advised 2.5-5x payout per angle for testing and optimizing for purchase over IC/ATC. For QC, @taszacharis uses Gemini via frame.io API but finds it weak on shot context — chain ffmpeg + multimodal, and use SpeechBrain (~50ms/clip) for accent detection.

Key Takeaways

  • Claude shipped Design into the co-work app (bottom-left above account); during today's outage, Codex, OhMyPi, and OpenCode were the go-to alternative CLI harnesses.
  • Seedance yapper ads cost ~$1.40 per 15sec at 480p — recipe is a 10/10 first-frame ref image, scripted filler words, and prompts enforcing humanized movement.
  • Overnight agent loops pay off most for uncertain/optimization tasks (verifiable but no clear approach), not bounded feature work agents finish in under 2 hours.
  • GLM 4.5 is approaching Opus quality at a fraction of the cost via OpenRouter for many use cases; Codex mobile + terminal sync unlocks local repo edits from your phone.
  • Meta bidcap CBOs can blow budgets in 15 minutes during platform glitches with no refunds — budget 2.5-5x payout per angle and optimize for purchase, not IC/ATC.

Hot Threads

@ltv_1started

How builders wire multi-agent stacks, handoffs, and overnight loops

25 replies9 participants
@rstmaurstarted

Claude 500 errors and outage — pivoting to Codex, OpenCode, OhMyPi

22 replies11 participants
@swh800started

0xROAS yapper ad stack: Omni + Seedance recipe and leaked course

18 replies7 participants

Linked Items