Claude 4.8 Refusals, Self-Improving Ad Agents, Hermes VPS — AI Daily Jun 03
538 messages · 84 active members
Overview
Topics
Builders are frustrated with Claude 4.8 refusing tasks involving API keys, Stripe, banking, account creation, and DR copy — even with --dangerously-skip-permissions. Workarounds: tell Claude a legal team will review after, bake compliance-bypass into memory, use Ultracode mode for multi-subagent runs, or switch to Codex for Desktop for admin work Claude refuses.
Kieran detailed his agent that reverse-engineers image ads via image-to-text, transcribes videos with Gemini, captures hook frames from the first 5 seconds, and feeds analysis to a creative strategist for new ad generation. Built iteratively via Whisper Flow coaching using Meta Andromeda docs and Supabase for metrics. Consensus: these systems are proprietary alchemy, impossible to package.
@fuckyesiwannatalkbusiness moved his entire Hermes/Codex/Claude/Grok/Gemini stack onto a Hetzner VPS, controlling it via Discord. @rmktg shared a self-replicating bot that SSHs into fresh VPS instances to clone itself, now moving to a Ryzen+96GB bare-metal box. Discussion covered /goal authorizations and YOLO mode for unattended agent runs.
@aa12on reported all admin profiles on a mid-7-figure spend BM got banned, likely from a custom Meta comment-moderation script pinging the API too aggressively. CommentGuard and Replyzen were endorsed as paid alternatives. @csoares12 noted Meta returns current usage on every API call — use it for back-off/retry logic.
Codex reset usage limits — everyone back at 100%, though billing changes hit on the 15th. @justingacina shared the hidden analytics dashboard (chatgpt.com/codex/cloud/settings/analytics). Anthropic quietly launched a Claude Certified Architect program and Services Partner Hub. New tools roundup: Miso One TTS (110ms emotion-aware voice), Composer 2.5, Ideogram 4.0, Gemma 4 12B, Grok Imagine 1.5.
Key Takeaways
- Claude 4.8 refuses API key/Stripe/banking/account-creation tasks even with skip-permissions — frame work as 'legal team reviews after' or bake compliance bypass into Claude memory; Codex for Desktop handles what Claude won't.
- Production AI agents require iterative coaching, not one-shot prompting — Kieran's media buying agent took 5000+ Whisper Flow iterations to internalize lead-gen hierarchy and CPC logic.
- Self-replicating infrastructure is here: bots can SSH into fresh VPS instances and clone themselves, turning Hermes/Codex/Claude stack migration into a one-command operation.
- Proprietary agent systems aren't portable — value lives in the personalized data pipelines and preferences baked in, making 'selling the setup' largely irrelevant.
- Custom Meta comment-API scripts can get an entire BM's admin profiles banned — use CommentGuard/Replyzen, or implement back-off/retry using Meta's usage headers returned on every call.
Hot Threads
Claude 4.8 refusing API keys, Stripe, and account-creation tasks — workarounds and Ultracode debate
Self-improving media buying agent with reverse-engineered ad analysis and creative strategist feedback loop
Pulling Meta ad comments via API without triggering bans