06-reference

alphasignal xai grok build skills claude md

2026-05-26·reference·source: AlphaSignal·by AlphaSignal
newslettercurationagentic-codingclaude-codegrok-buildharness-engineeringagent-skillssupply-chain-securityai-news

AlphaSignal 2026-05-26 — Grok Build now reads CLAUDE.md + supports Skills (the skills question is answered)

Why this is in the vault

The lead item resolves the open question we logged on 2026-05-15: when Grok Build first shipped, AlphaSignal copy didn't say whether it had a Skills surface, so we filed "Skills/plugins ecosystem is the unknown" as a real evaluation item. Today's issue answers it directly — Grok Build extends via Skills, hooks, and MCP servers, and reads your existing CLAUDE.md with no changes needed. That last clause is the load-bearing detail: xAI is explicitly courting the Claude Code harness ecosystem rather than building a walled garden, which changes the competitor/complement framing for the existing Notion eval task. The rest of the issue is routine curation; one secondary item (Perplexity Bumblebee, an MCP-config supply-chain scanner) lightly touches RDCO's no-secrets-on-disk / MCP-install-security posture.

⚠️ Sponsorship

This issue carries three paid placements plus the standing "Work With Us" CTA. Disclosed so the curation isn't read as neutral editorial:

Issue contents

Editor's intro frames the day as a terminal land-grab: "The terminal is the new battleground... every major lab now wants to live in your shell," with the new feature checklist being Plan Mode, parallel subagents, CI pipelines. Author byline: Lior Alexander (AlphaSignal founder). Three "Top" features (one Top News + two Top Repo) + six Signals + three sponsor blocks.

Top News — xAI opens Grok Build beta (5,596 likes). Terminal-native coding agent that reads the project, makes a plan, edits files, runs commands. Pitched against the "starts executing and goes off the rails" failure mode via Plan Mode (shows the full plan, you approve each step, nothing touched until you say so). New/confirmed capabilities this issue:

Top Repo — Perplexity open-sources Bumblebee (3,292 likes). Read-only supply-chain security scanner Perplexity uses internally. Context: one attack group injected malicious code into 160+ packages including a React tool at ~12M weekly downloads. Bumblebee scans browser extensions (Chrome/Edge/Brave/Arc/Firefox), editor plugins (VS Code + forks), MCP config files (the local settings that tell AI assistants which external services they can reach — email, DBs, repos), and vulnerable packages across npm/PyPI/Go. Read-only by design so it can't trigger the malicious code it's scanning. Apache 2.0, free on GitHub.

Top Repo — ZOZO open-sources ppf-contact-solver (2,751 likes). GPU physics solver for cloth/rope/soft-body contact: 180M+ contact points per scene, no clipping, strict stretch limits. Python API + JupyterLab, Docker (~1GB), Windows .exe, cloud-ready (AWS/GCP/RunPod), community Blender add-ons. Runnable on a sub-$0.50/hr vast.ai GPU.

Signals.

  1. Pliny jailbreaks Qwen 27B to a 4% refusal rate with zero capability loss (2,445 likes).
  2. (sponsor) Brave Search API for RAG — see Sponsorship above.
  3. New benchmark: coding agents lose ~30% pass rate when real databases are added (498 likes).
  4. LongCat ships a free MIT-licensed talking-avatar model "that may beat all others" (1,725 likes).
  5. New local coding model beats Qwen and DeepSeek on 128 GB RAM machines (777 likes).
  6. Google Antigravity ships a terminal CLI to run AI agents from the command line (1,386 likes).

Mapping against Ray Data Co

Strong — Grok Build's Skills + CLAUDE.md compatibility directly updates the Notion eval task. This is the item that justifies the note. The existing board task "Evaluate xAI Grok Build CLI as Claude Code competitor / complement" (Notion id 361f7d49-36d1-8198-a8e0-cd5e15c0b5a3, opened off the 2026-05-15 capture) explicitly listed the Skills/extensibility surface as the unknown that gated the competitor-vs-complement decision. That unknown is now resolved in xAI's favor for adoption ease: Grok Build supports Skills, hooks, and MCP, and reads our existing CLAUDE.md unchanged. Three implications for RDCO's harness/skills posture:

  1. The harness layer is now genuinely contested, and xAI chose interop over lock-in. RDCO's entire architecture is a fat-skills bet on the Claude Code harness — ~/.claude/skills/ directory, /improve self-edit loop, the verify-* critic gates, CLAUDE.md as the rule substrate. Grok Build reading CLAUDE.md unchanged means RDCO's configuration capital (the hard rules, precedence chain, ~100 skills) is no longer Claude-Code-exclusive. The skills/config layer is portable in principle; the moat is the harness quality and the model behind it, not the file format. This is consistent with the SkillOpt finding that optimized skills transfer Codex ↔ Claude Code ([[2026-05-26-skillopt-self-evolving-agent-skills]]) — skills are increasingly a cross-harness asset, which lowers switching cost in both directions.

  2. Reframes the eval task from "competitor" toward "complement / portability hedge." Because Grok Build ingests CLAUDE.md and Skills with no rewrite, the cheap experiment is no longer "rebuild RDCO on Grok" — it's "point Grok Build at the existing ~/.claude/ config and diff the behavior." That's a low-cost portability test worth adding as a concrete sub-step on the Notion task: does our precedence chain, hard-rule set, and a representative skill (e.g. a verify-* critic) actually behave under Grok Build, or does "reads CLAUDE.md" mean parses-but-ignores-semantics? Skills/hooks/MCP parity collides directly with Anthropic's skills-marketplace story ([[2026-04-27-alphasignal-anthropic-claude-marketplace-agent-quality]]); if Grok Build's skill execution is faithful, Anthropic no longer owns the extensibility narrative outright.

  3. Plan Mode = our review-before-execute pattern; parallel subagents = our pipeline pattern. The two features AlphaSignal leads with are the exact control-flow primitives RDCO already runs (Plan-Mode-style approval gates, the 4-seat pipeline + per-question subagent fan-out). That's reassurance the architecture bet is industry-validated, not idiosyncratic — but it also means RDCO's differentiation can't be those primitives. The edge has to be the accumulated skill library + verify-gate discipline + the COO-agent's domain config, which (per implication 1) is now partially portable. Net: the model and harness-execution quality become the real comparison axes, not feature checkboxes.

Weak/background — Perplexity Bumblebee touches the MCP-install-security posture. Bumblebee scans MCP config files for poisoned entries, which is adjacent to RDCO's standing rules: no-secrets-on-disk (1Password wrappers, never .env) and the mandatory security-review-before-MCP/plugin/skill-install SOP. Bumblebee is a read-only config scanner, not a replacement for the install-time review, but it's a candidate tool to evaluate as an ongoing MCP-config integrity check (vs. the current point-in-time pre-install review). Low priority; note only.

Skip — ZOZO physics solver, LongCat avatar, Qwen jailbreak, local 128GB coding model, Antigravity CLI. No RDCO surface. (Antigravity's terminal CLI is more terminal-land-grab evidence but adds nothing actionable over the Grok Build read.)

Curation section — notes

Related