Kimi K2.6 Orchestration, Opus 4.8 Burn, Claude Skills — AI Daily May 30

363 messages · 64 active members

363
messages
64
active members
@jcartu, @arielletolome, @samb69
top contributors

Overview

Orchestration strategy dominated the day: @jcartu made the case for routing Hermes through Kimi K2.6 Turbo on Fireworks (200+ tps, ~4x cheaper than Opus) and escalating to Opus 4.8 only for hard reasoning. He shared two ASCII architecture diagrams (green/yellow/red task lanes and a four-layer orchestrator/reasoner/bulk stack) that @rmktg and @samb69 started implementing. @yangthegoat confirmed Anthropic killed Claude sub auth inside Hermes, forcing the move to API-based providers like Fireworks and OpenAI. Opus 4.8 sentiment kept sliding — @zippi101 reported 30-minute single-prompt waits, @mb29266 said 4.6 still beats 4.8 for structured ad copy and burns way fewer tokens, and @samtome argued 4.8 is basically 4.7 with workflows. Builders also unpacked Claude's $20/$200 Max tiers (1:1 API credits for claude -p, but exhausted fast under agent loads) and tips for persisting learnings via claude.md, memory.md, and explicit Skills. @BowTiedRabbit's drop of Matt Pocock's open-source .claude skills repo spread fast, with @WhiskeyATX extending /improve-codebase-architecture into a sub-agent that reviews PRDs. Infra and shipping news rounded out the day: Coolify surfaced as a self-hostable Vercel/Railway alternative with GitHub webhook auto-deploy; Paulo Hernandez detailed a 52-agent Sky House orchestration platform with a 7-agent review team that must pass two clean rounds before code ships; @arielletolome scraped 9.9k Newsbreak ads (50k FB ads incoming) for an auto-angle creative brief system; @jonmacofficial is launching ViralView Monday (BYO Kie.ai, Kling 3.0 over Seedance); and Grok Imagine 1.5 hit the X API with strong visuals but weak audio at ~$0.08/sec.

Topics

@jcartu shared a tiered orchestration pattern using Kimi K2.6 Turbo on Fireworks (200+ tps, 256k ctx) as the default orchestrator, Opus 4.8 for high-stakes reasoning, and DeepSeek V4 Flash for bulk work, with two ASCII diagrams and ~3x cost savings vs gpt-5.5. @rmktg and @samb69 are migrating; @yangthegoat confirmed Anthropic killed Claude sub auth inside Hermes, forcing API providers from day one.

@zippi101 sees 30-min waits on single Opus 4.8 prompts, @kennetele and @mb29266 say 4.6 still wins for structured ad copy and burns fewer tokens, and @samtome argued 4.8 is 4.7 with workflows. Builders also mapped Claude's $20/$200 Max tiers to 1:1 API credits for claude -p but found burn rates steep, eyeing GPT Pro and Grok, and sharing claude.md/memory.md persistence tips.

@BowTiedRabbit shared github.com/mattpocock/skills and called grill-with-docs a game changer for new features. @WhiskeyATX dropped it into a live project and extended /improve-codebase-architecture as a sub-agent that reviews PRDs from an architectural lens. @jonmacofficial built a custom skill from the workflow; @iannagy installed immediately.

Coolify emerged as a self-hostable open-source alternative to Vercel/Heroku/Railway with 280+ one-click services and GitHub webhook auto-deploy, validated by members on Oracle free tier and VPS. Paulo Hernandez detailed Sky House — a 52-agent platform organized like a company with a top-level brain delegating to department agents, plus a PRD-to-subagent flow requiring a 7-agent review team to pass two clean rounds before shipping.

@arielletolome scraped 9.9k Newsbreak ads (50k FB ads incoming, ~90 hrs) for a creative brief system that auto-generates 20 ad angles and image briefs at a time. @jonmacofficial launches ViralView Monday (BYO Kie.ai, Kling 3.0 over Seedance). Grok Imagine 1.5 hit X API (~$0.08/sec) with strong visuals but weak audio/VO; Seedance remains the community pick. @expadz shared a Claude Code video editing workflow that trims silences and rerolls inconsistent clips.

Key Takeaways

  • Don't drive Hermes with Opus directly — route through Kimi K2.6 Turbo on Fireworks (~200 tps, ~4x cheaper) and escalate to Opus 4.8 only for red-lane reasoning; Anthropic killed Claude sub auth inside Hermes.
  • Opus 4.8 burns more tokens and runs slower than 4.6 for many real workflows; 4.6 still wins for structured ad copy, and Claude's $20/$200 Max API credits exhaust fast under agent loads.
  • Install Matt Pocock's .claude skills repo — grill-with-docs for new features and /improve-codebase-architecture (runnable as a sub-agent to review PRDs) are immediate wins; after feedback, tell Claude to 'update the learnings to github' to auto-persist claude.md and memory.md.
  • Coolify gives you GitHub-webhook auto-deploy on your own VPS as a Vercel/Railway replacement, and multi-agent PRD review (7-agent team, two clean rounds) catches weaknesses before code is written.
  • Grok Imagine 1.5 has strong visuals but weak audio at ~$0.08/sec; Seedance still wins for video, and ViralView (BYO Kie.ai, Kling 3.0) launches Monday for cloning viral TikToks/Meta ads.

Hot Threads

@jcartustarted

Kimi K2.6 Turbo orchestrator stack with ASCII architecture diagrams replacing Opus for Hermes

18 replies7 participants
@samb69started

Grok Imagine 1.5 video testing, Seedance comparison, and Coolify self-hosting discovery

20 replies6 participants
@arielletolomestarted

Scraping 50k FB ads to power an auto-angle creative brief system

12 replies6 participants

Linked Items