Claude Code at 900k LOC, Gemini 3.5 Flash, Ad Workflows — AI Daily May 21

519 messages · 83 active members

519
messages
83
active members
@samb69, @jasonakatiff, @mb29266
top contributors

Overview

Today's headline thread came from @jasonakatiff hitting the ceiling of Claude Code at ~900k lines — billing, subscriptions, and routing logic drifting as Claude forgets 10-20% of touch points per change. The community debated SWE audits vs. multi-loop discovery and Codex-driven deployment reviews. Consensus: vibe-coded systems work surprisingly well at revenue scale, but unified domain boundaries and orchestrator-grade context are the missing piece. On infra, @instanetworks cut Neon compute bills 66% by just asking Claude to audit. On the model front, Google dropped Gemini 3.5 Flash plus Antigravity 2.0, an agent-first desktop IDE that reportedly built an OS from scratch in internal tests. SpaceX's IPO filing revealed Anthropic is paying them $1.25B/month ($15B/yr) for compute. A GitHub breach (3,800 repos via a poisoned VS Code extension) added urgency to agent security discussions, where @rstmaur and @seekersight shared layered review stacks (Claude + Codex adversarial + Greptile, pre-push hooks, scoped credentials, Tailscale gateways). The copywriting/ad track ran hot: @nickistage39 vented about robotic Claude copy for a telederm brand, leading to a deep workflow thread — run the same research across Perplexity, Gemini, Claude, GPT, and Grok 4.3, then extract the best. @mb29266 described launching 300 ads per session via Claude-generated long copy + matching images. @rockdm asked if anyone has fully replaced a creative strategist with AI (answer: not yet — agents scale good ideas but can't originate them). An arxiv paper claiming rude prompts get ~5% better accuracy sparked a side debate on whether politeness is suboptimal.

Topics

@jasonakatiff shared that Claude Code is struggling to track dependencies across his ~900k LOC platform — billing, routing, and inventory logic drifting because the model only considers ~80% of touched surfaces per change. The thread debated formal SWE audits of core paths, multi-loop discovery, and Codex audits of the last 20 deployments. Key insight: unify duplicate flows (one-time vs subscription billing) and enforce hard domain boundaries. Bonus: @instanetworks cut Neon compute 66% via a Claude-driven audit.

Google released Gemini 3.5 Flash, positioned as their most powerful coding/agent model, alongside Antigravity 2.0 — a standalone desktop IDE built as the model's native runtime. Internal tests reportedly had it building an OS from scratch. Some users are already reporting flicker glitches in Antigravity. Adjacent news: SpaceX's IPO filing shows Anthropic pays them $15B/yr for compute.

@nickistage39's robotic Claude copy sparked a deep workflow thread: run the same research across Perplexity, Gemini, Claude, GPT, and Grok 4.3, then extract the best from each (Grok 4.3 won for teen brand voice). Load dozens of real examples instead of instructions, use speech-to-text to ramble vibes. @mb29266 described launching 300 ads per session via Claude long copy + matching images. @rockdm's question on replacing a creative strategist landed at: agents scale human-curated bangers, they don't originate them.

@seekersight shared his enforcement stack: Claude reviews design doc vs codebase, Codex does adversarial review for auth/race conditions/CSRF, Greptile catches cross-file drift, with pre-push hooks and a 2-pass Codex cap to avoid infinite fix loops. @rstmaur added safe-mode-first, separate agent lanes, scoped credentials, Tailscale gateways, and treating web/Telegram/docs as untrusted data. The GitHub breach (3,800 repos via poisoned VS Code extension) underlined urgency.

@samb69 argued people already making $1M/mo will 10-20x with AI while normies stay flat — pointing to viral footage of college students booing Eric Schmidt. @A1a5h3 cited Toffler: the new illiterates are those who can't learn, unlearn, and relearn. A side debate on whether rude prompts actually get 5% more accurate Claude output (per arxiv) reinforced that defaults aren't optimal.

Key Takeaways

  • Claude Code remains viable past 900k LOC for revenue products, but enforce domain boundaries and multi-loop discovery to catch the 10-20% of surfaces it misses per change.
  • Ask Claude/Codex to audit your cloud bills — @instanetworks cut Neon compute 66% with no perf hit, just by asking for optimization suggestions.
  • For copywriting, drop instructions and load examples — run the same research across Perplexity, Gemini, Claude, GPT, and Grok 4.3, then extract the best (Grok 4.3 won for teen brand voice).
  • Replacing a creative strategist with AI fails without a seeded bank of human-curated winning ideas; agents scale ideas, they don't originate good ones.
  • Strict gating (pre-push hooks, 2-pass Codex caps, scoped credentials, treating external data as untrusted) is what makes multi-agent workflows safe for prod — especially after the 3,800-repo GitHub breach.

Hot Threads

@jasonakatiffstarted

Hitting Claude Code's ceiling at 900k LOC — when to bring in a SWE

35 replies6 participants
@nickistage39started

Getting Claude to write usable copy for a telederm acne brand

28 replies8 participants
@rockdmstarted

Has anyone actually replaced their creative strategist with AI?

14 replies4 participants

Linked Items