06-reference

alphasignal ultraplan karpathy claude

Sun Apr 12 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: AlphaSignal ·by Lior Alexander

“Anthropic Ultraplan: Cloud Planning + Local Execution for Claude” — AlphaSignal

Why this is in the vault

Three items directly relevant to RDCO: (1) Anthropic’s Ultraplan separates planning from execution in Claude Code, which may reshape our agent architecture; (2) a Karpathy-inspired CLAUDE.md ruleset for constraining LLM coding behavior, directly comparable to our SOUL.md + CLAUDE.md approach; (3) Claude Cowork enterprise controls signal Anthropic’s push into the governance layer we rely on.

Curation section — notes

  1. Anthropic Ultraplan — New Claude Code feature that offloads implementation planning to a cloud/browser session while keeping execution local. You run /ultraplan <task>, review a structured plan in-browser with inline comments, then choose to execute in cloud or send the plan back to terminal. Solves the problem of reviewing long plans in a terminal-only workflow.

  2. Karpathy-inspired CLAUDE.md ruleset (20K+ GitHub stars) — A single CLAUDE.md file defining four rules to constrain LLM coding: state assumptions before acting, restrict to minimal implementations, limit changes to requested scope, require explicit validation before completion. Targets failure patterns Karpathy identified: silent assumptions, overbuilt code, scope creep.

  3. Google PaperOrchestra — Multi-agent framework converting raw research notes into LaTeX manuscripts. Five specialized agents (outline, literature, plotting, writing, refinement). Evaluated on 200 CVPR/ICLR 2025 papers; claims 50-68% win margin on literature review quality over baselines.

  4. Claude Cowork enterprise controls — Anthropic adding governance and observability tooling for organization-wide Cowork deployment.

  5. Signals — MiniMax M2.7 open-weight model; Google adding interactive simulations to Gemini; Replit deploying to Databricks; OpenClaw stability updates.

Bias and sponsor notes

Bright Data and Innovating With AI are paid sponsors. Neither sponsor’s product overlaps with the editorial items. No detectable editorial bias from sponsorship. AlphaSignal’s curation is engagement-metric-driven (likes/stars counts featured prominently), which skews toward popularity over substance.

RDCO mapping

Karpathy’s CLAUDE.md rules vs. our SOUL.md + CLAUDE.md

The four rules (clarify assumptions, minimal code, scope discipline, explicit validation) are a subset of what we already enforce. Our SOUL.md goes further by defining personality, decision authority, and communication norms — not just coding constraints. The cross-check from April 12 flagged that our CLAUDE.md is growing and may need pruning; Karpathy’s approach of a concise, focused ruleset reinforces the resolver audit recommendation to keep instruction tokens under control.

Action: Compare our current CLAUDE.md token count against the Karpathy four-rule baseline. If ours exceeds the Tan 200-line benchmark, identify what can move to on-demand skill loading.

Anthropic Ultraplan and our agent architecture

Ultraplan introduces a cloud-local split for planning vs. execution. This is architecturally significant for RDCO:

Watch item: Monitor whether Ultraplan supports programmatic invocation (not just /ultraplan in CLI). If it exposes an API, it could be integrated into our autonomous loop as a planning step.

PaperOrchestra — multi-agent writing pattern

The five-agent pipeline (outline, literature, plotting, writing, refinement) mirrors patterns we’ve discussed for content production. The “simulated peer review” refinement agent is the same concept as our harness review and draft-review skill. Worth tracking as a reference architecture for structured multi-agent content workflows.

Cowork enterprise controls

Anthropic adding governance and observability to Cowork aligns with Jaya Gupta’s moat thesis — permission and trust as the scarce asset. Enterprise controls make Cowork viable for consulting clients who need audit trails.