Loop Engineering, Codex vs Claude, WWDC Flop — AI Daily Jun 08

617 messages · 90 active members

617
messages
90
active members
@jcartu, @rmktg, @jasonakatiff
top contributors

Overview

Sunday's chat was anchored by a clear thesis shift: prompt engineering is giving way to loop engineering. @seekersight built an agentic loop using vision.md + steward.md, risk scoping, automated tests, and approval gates — waking up to a working prototype. The community swapped Ralph-loop tactics, /goal command patterns with embedded authorizations, and pointed newcomers to Steve Yegge, Geoffrey Huntley, and a leaked Google playbook for AI coding agents. A viral example: a developer automated an entire job with Claude loops in 24 minutes. On the tooling front, Codex (ultracode/xhigh) gained serious ground against Claude 4.8, with @jasonakatiff and others fully migrating to Codex in cmux citing 4.8's chattiness, bugs, and overcomplication. A rumored Anthropic 'Fable' (diet Mythos) model is expected to drop within a day. Meanwhile @jcartu shipped 3D Tetris in ~8 minutes on local DeepSeek FP8 at 200 tps, arguing 2x RTX 6000 Pros can sustain a 5-person dev team with 384k context. Apple's WWDC keynote was widely panned as a 65-minute Siri-meets-Gemini snoozefest with no Mac mini/Studio refresh — stock dropped ~2.5% during the stream. Ad workflows kept evolving (Seedance 2.0, Kling 3, Kie.ai at ~$6/45sec), Meta ad rejection throttling drew consensus tactics, and builders increasingly replaced SaaS with vibe-coded systems — one member saving $2k/mo. As @Anonymoushat noted from an SF 24/7 cafe with 30+ Claude Code screens running: 'everyone's talking MRR, no one's talking profitability.'

Topics

@seekersight detailed a two-file orchestration pattern: vision.md as north star (what good looks like, risks, approval gates) and steward.md as the operating layer (morning docket, priorities, approval queue, recap). Combined with Ralph loops, /goal commands with embedded authorizations, and a viral 24-minute job automation via Claude loops, the frontier is shifting from one-off prompts to designed agent loops with clear autonomy boundaries.

Active debate on Claude 4.8 — praised by some but criticized as too chatty, overcomplicating tasks, and buggy in CC. @jasonakatiff downgraded from $200 to $20 plan and moved fully to Codex in cmux; others report ultracode handling big prompts at ~10% of weekly Claude quota. Codex's terse output and desktop app are now features, not limitations. Rumored 'Fable' (diet Mythos) Anthropic model expected within a day.

The group live-roasted Apple's 65-minute keynote — mostly Siri hooking up to Gemini, no Mac mini/Studio refresh, and a ~2.5% stock drop (~$112B) during the stream. M5 Mac mini now pushed to fall 2026. Members debated whether Apple is structurally behind in the AI race given iPhone revenue exposure to AI-native phones and foldable competition.

Seedance 2.0, Seedance Fast, and Kling 3.0 compared for UGC ad generation, with @arielletolome moving from Runway to Kie.ai at ~$6/45sec. @rutty123's Seedance 2.0 character swap demo wowed the group. @jasonakatiff shared his multi-7-fig pet supplement brand workflow: foreplay MCP for competitor ads → Gemini breakdown → Claude skills for scripts. On Meta: never delete rejected ads — swap creative, get approval, then disable.

@jcartu shipped 3D Tetris in ~8 minutes on local DS4f FP8 at 200 tps, claiming 2x RTX 6000 Pros (~$24k) support a 5-person team 24/7 with 384k context (1M with 4 GPUs). Parallel theme: builders replacing paid SaaS with Claude Code-built systems — @nickistage39 cutting $2k/mo, @samb69 migrating to Herdr from cmux+tmux for cleaner TUI and Moshi mobile support.

Key Takeaways

  • Loop engineering > prompt engineering: give agents vision.md + steward.md, risk tiers, test loops, and approval gates — humans only for prod/secrets/customer data
  • Power users are migrating from Claude 4.8 to Codex ultracode citing bugs, verbosity, and overcomplication — watch for 'Fable' model drop within a day
  • Apple's WWDC delivered no new Mac mini/Studio (M5 pushed to fall 2026); stock dropped 2.5% during a Siri+Gemini-heavy keynote
  • Never delete rejected Meta ads — swap to compliant creative, wait for approval, then disable; Kie.ai is winning Seedance 2.0 hosting at ~$6/45sec
  • Local DeepSeek FP8 at 200 tps on 2x RTX 6000 Pro (~$24k) can sustain a 5-person dev team — local iron increasingly viable vs cloud APIs

Hot Threads

@rmktgstarted

Live-reacting to Apple WWDC keynote disaster (Siri+Gemini, no hardware)

35 replies10 participants
@seekersightstarted

Loop engineering agent with vision.md, steward.md, and overnight prototypes

22 replies9 participants
@jasonakatiffstarted

Codex xhigh/ultracode vs Claude 4.8 — full migration to cmux

14 replies6 participants

Linked Items