Opus 5 Prompt Cuts, Anthropic Outages, Kimi K3 — AI Daily Jul 29

434 messages · 75 active members

434
messages
75
active members
@jcartu, @samb69, @samtome
top contributors

Overview

Prompt engineering evolution and platform reliability defined the day. @iamgalba showed that trimming ~70% of verbatim from CLAUDE.md, skills, and progressively-disclosed docs made Opus 5 dramatically faster and better — echoing Anthropic's own 80% system prompt reduction. Consensus: verbosity accumulated as scaffolding for weaker models now actively hurts newer ones. That lesson was tested under fire as multiple waves of Anthropic outages hit Claude Code, with Opus 5 the slowest to recover; builders with auto-failover to GLM 5.2 and Kimi K3 (often via Headroom proxy caching, which also saves 15-50% tokens) stayed unblocked while others fell back to Sonnet. Model comparisons continued with extensive Kimi K3 testing — @samtome rated it a solid self-correcting workhorse hampered by a buggy native harness, while @jcartu found it relentless when run inside OMP with Fireworks. Fable, by contrast, got widely panned as degraded this week, behaving 'like Opus 4' and losing context within minutes. Orchestration model choice split the room: Opus 5 everywhere vs. medium-for-orchestration with delegated deep-review workers. On tooling, Orca is displacing cmux for agent workflows (pair with Claude Code remote-control or Tailscale for mobile), OMP's forced role→model routing was defended as guardrail hygiene against runaway subagent spawns, and a rich thread emerged on converting books (Cialdini, Ariely, Bencivenga, Psycho-Cybernetics) into reusable Claude skills rather than dumping full texts into context. Compliance discussion covered Meta pixel + medical intake as a HIPAA no-go, and Meta's Q2 miss with $145B AI capex guidance raised sustainability questions.

Topics

Opus 5 Verbosity Cuts Boost Performance18 msgs

@iamgalba trimmed ~70% of instructions from CLAUDE.md, skills, and progressively disclosed docs and saw Opus 5 run faster and better with reduced token consumption. Anthropic reportedly cut Claude Code's own system prompt by 80%. The old verbosity was reasoning-scaffolding for weaker models and now hurts newer ones — give the agent more space to think.

Anthropic Outages and Failover Strategies25 msgs

Multiple waves of Claude Code errors hit throughout the day, with Opus 5 the slowest to recover. Builders reported success switching to Sonnet mid-session, and auto-failover to GLM 5.2 and Kimi K3 held up for those with it configured. Headroom proxy caching was recommended as both a resilience layer and a 15-50% token saver.

Kimi K3 Impressions and Harness Bugs25 msgs

@samtome rated K3 a solid ~4.8-level workhorse that occasionally surprised with mid-trace self-correction, but flagged awful harness UX: 422 errors, no retry on rate limits, false-completed todos, desktop glitches. @jcartu counters that K3 in OMP paired with Fireworks is the most relentless model he's tested. Fable, meanwhile, is getting panned this week as regressed to Opus 4-level behavior.

Book-to-Skill Conversion for Claude15 msgs

Users are converting books into Claude skills instead of dumping full texts into context. Workflows include chapter-by-chapter sessions (Psycho-Cybernetics example) and building persuasion stacks from Cialdini's Influence, Pre-Suasion, Ariely's Predictably Irrational, and Bencivenga's 100 Seminar. Shared drives with source PDFs circulated.

Orca Replacing cmux for Agent Workflows15 msgs

Several builders reported going 100% on Orca and abandoning cmux, though mobile output truncation and clunky UI remain pain points. Orca supports Claude Code remote-control and Tailscale connections, making mobile session management viable. Some still mix cmux and herdr into Orca's workspace terminal. Related: OMP's forced role→model routing was defended as guardrail against runaway subagent token burn.

Key Takeaways

  • Aggressively trim CLAUDE.md and skill docs — Opus 5 (and Anthropic's own prompts) perform better with 70-80% less verbosity and more agent freedom.
  • Two Anthropic outages hit today — auto-failover to GLM 5.2 or Kimi K3 (ideally via Headroom proxy, which also cuts tokens 15-50%) kept builders unblocked.
  • Kimi K3 is smart and self-correcting but its native harness is buggy; run it in OMP or via Fireworks for a better experience.
  • Convert books into Claude skills instead of loading full texts — persuasion stacks (Cialdini, Ariely, Bencivenga) are the popular starting point.
  • Lock agents to specific models in multi-model harnesses — free-choice routing risks runaway token burn from over-eager subagent spawning.

Hot Threads

@momodxbstarted

Anthropic elevated errors and failover to GLM/Kimi via Headroom

15 replies8 participants
@samtomestarted

Kimi K3 raw harness testing — smart model, painful UX

18 replies5 participants
@iamgalbastarted

Cutting 70% of prompt verbosity makes Opus 5 dramatically better

12 replies4 participants

Linked Items

vLLM on X

x.com
@samtome

OpenAI on X

x.com
@mikeconner

GitHub - openai/codex-security: SDKs and CLI for Codex Security · GitHub

github.com
@mikeconner

GitHub - digimata/quill: Ultraminimalist macOS recording + transcription. · GitHub

github.com
@Wootbro

GitHub - murtaza-nasir/speakr: Speakr is a personal, self-hosted web application designed for transcribing audio recordings · GitHub

github.com
@leewardbound