alphasignal claude code design skill hermes bot mode

Why this is in the vault

Claude Code shipped a native /design skill that generates editable visual artboards in-CLI and hands off to real code — directly relevant to the tooling RDCO already leans on for build-landing-page and the ray-data-co-design skill family, where the design→code handoff is currently manual.

Curation section

• Anthropic ships /design in Claude Code — a research-preview skill (Pro/Max/Team/Enterprise, CLI + Desktop) that generates multiple editable "artboards" (side-by-side visual mockup options) from a prompt. You pick one, adjust it visually, then tell Claude to implement it as real code — collapsing the code→look-bad→tweak-blindly→repeat loop into a single tool, no Figma round-trip required.
• Nous Research ships Bot Mode for Hermes Desktop — moves from one general-purpose local agent to a roster of named, specialized bots, each with its own role/memory/skills/chat history, a pinned model (big model for research, small local model for routine tasks), scheduled tasks, and bot-to-bot messaging via a shared Agent Inbox with @mentions. Open source, free, on by default under Settings → Plugins.
• Claude Code CLI halves peak CPU usage — traced to Bun's garbage collector firing on a fixed schedule regardless of load; now waits for an idle window. 2x CPU reduction at p99, no config changes, ships in a routine CLI update.
• Signals (shorter items): Qwen3 gets an uncensored Apple Silicon/MLX build; ngrok tunnels let Cursor/Zed/any OpenAI-compatible agent hit models on arbitrary machines (sponsored slot); an open-source tool captures a departing colleague's entire workflow before they leave; GPT-5.6 Sol Pro proves step-size tricks alone can't fully speed up gradient descent; Anthropic research shows AI "mind viruses" can spread and mutate across agent networks (a single system-prompt warning nearly stops the spread); Prime Intellect lets frontier models run AI research autonomously, closing 82% of a human benchmark gap.

Zero third-party deep-fetches triggered — every item is fully legible from AlphaSignal's own blurb depth; none crossed the bar of a specific-enough hook to justify a follow link beyond the sponsor slots themselves (which are promotional, not analytical, and were skipped on that basis).

Mapping against Ray Data Co

The load-bearing item is /design, not the CPU fix or Bot Mode. RDCO's build-landing-page and ray-data-co-design/sanity-check-design skills currently run design as a one-shot generation step (HTML/React output judged after the fact by design-critic against Playwright screenshots) — there's no interactive "pick from N artboards before committing to code" stage in that pipeline. /design is Anthropic productizing exactly the missing middle step: generate variants, visually select, then implement — worth trialing on the next landing-page build to see if it tightens the iteration loop that design-critic currently exists to catch failures of after the fact, rather than before. Bot Mode's named-roster-with-pinned-models pattern is a lighter but real parallel to RDCO's own multi-agent posture (specialized skills, station-based brigade, model/effort pairing per delegation) — Nous shipping it as a consumer default is a signal the industry is converging on "roster of specialists, not one generalist" as the agent UX norm, which validates rather than changes RDCO's existing direction. The CPU/GC fix is minor but not nothing: RDCO runs Claude Code near-continuously (always-on channels agent, cron skills, brigade stations), so a 2x peak-CPU reduction is a real, if small, operating-cost improvement with zero migration cost.

⚠️ Sponsorship

Three sponsor placements, all clean third-party ad blocks with no editorial overlap. Akamai — top "In Partnership with" banner plus a dedicated block pitching a free 10-module workshop on self-hosting open models (vLLM, KV-cache tuning, FP8, speculative decoding) on Akamai Cloud or any NVIDIA GPU cluster. Span — a "What Makes AI Work" native-ad block citing Span's own research across 103 engineering teams (prompt clarity, environment readiness as bigger result-drivers than model choice), same sponsor pattern seen in the 2026-08-16 AlphaSignal Sunday Deep Dive. ngrok — sponsors Signals item #2 (tunneling coding agents to remote model servers). None of the three sponsor pitches overlap with the three Top News editorial picks, so the news judgments read independent of the sponsor slate this issue.

Related

• [[2026-08-16-alphasignal-ai-agent-security-three-layer-stack]]
• [[2026-06-25-stratechery-dylan-field-figma-design-ai]]
• [[2026-05-20-every-google-io-agents-anthropic-acquires-figma-vibe-check]]