Claude 4.7 Backlash, Meta Ad Bans, $1 Adsets — AI Daily May 03
619 messages · 68 active members
Overview
Topics
Builders are abandoning Claude 4.7 en masse, citing slowness, refusals ('time to go to sleep' lectures), and confidently broken output. @jonmacofficial, @namtalks, and @jrizzolo report Codex xhigh and GPT 5.5 outperforming Opus on hard bugs with far better usage limits. The emerging pattern: plan in Claude, execute in Codex, or downgrade to 4.6 high. @samb69 reported positive early results on Opus 4.7 as a contrarian data point.
@mb29266 and @drcopybymatt shared production setups: private Meta dev apps with system user tokens shared to BMs, batched 50-ad uploads, portable creative drafts that strip page/pixel/optimization fields so adsets move freely across farmed BMs, and 2-minute adset delays to dodge automation flags. Consensus on bans is they're effectively random — keep multiple admins, BMs, and ad accounts ready, and protect your pixel on a safe BM. System user tokens may reduce profile ban risk vs user tokens.
A strategy of launching huge volumes of $1.28 single-ad adsets — popularized by Brazilian media buyers on YouTube — was tested and validated by several members, with reports of significantly lower CPMs/CPCs. Winners get promoted into 1-50-1 then 1-250-1 prescale campaigns. Main constraint is needing many pages and ad accounts to handle 250-ad page limits.
@weslindquist runs 3 Claude Code + 8 OpenClaw + 1 Hermes orchestrator agent over Slack — CC is more precise but OC tries to one-shot everything. @jasonakatiff burned a weekend trying to harden Claude Code env var handling (ultimately reverting to .env.local), while @expadz shared a 1Password CLI workflow with prod secrets in Railway/Vercel. Jason also flagged prompt injection from researched websites as a real exfiltration vector; @navuud argued frontier models defend well, and most builders prefer scoped admin API endpoints over raw DB service-role keys for agents.
@jcartu detailed cutting Anthropic spend from €6-7k/mo to €200/mo by running Qwen 3 Coder Next and Gemma 27B on €16k of home GPUs via official Nvidia vLLM containers for production-grade reliability. He also flagged OpenCode + GLM 4.7 on Cerebras hitting 1500 tps for cheap auditable bulk coding. Pair with Opus for planning and auditing.
Key Takeaways
- Claude 4.7 is regressing for many builders — migrate to Codex (medium for speed, high/xhigh for hard problems) or fall back to 4.6 high while Anthropic sorts it out.
- Meta bans are effectively random — redundancy in profiles, BMs, and ad accounts matters more than 'clean' behavior; system user tokens via private dev apps reduce surface area.
- The Brazilian $1.28 adset mass-launch strategy is producing notably lower CPMs and feeding 1-50-1 / 1-250-1 prescale ladders — but requires page and account inventory.
- Hermes is emerging as a viable orchestrator above Claude Code and OpenClaw fleets; communicate via Slack/Telegram and treat prompt injection from researched sites as a real exfil vector.
- Local inference (Qwen 3 Coder Next, GLM 4.7, Gemma 27B) on official Nvidia vLLM containers can replace 90%+ of Anthropic spend if you keep Opus for planning and auditing.
Hot Threads
$1 adset strategy results and 1-50-1 / 1-250-1 scaling mechanics
Portable Meta ad drafts across farmed BMs with system user tokens
Claude 4.7 is broken — cancelling and moving to Codex