Anthropic Just Dropped Ultra Plan for Claude Code

Watch on YouTube

📖 Description

Ray Amjad reviews Claude Code’s new /ultraplan command — a feature that offloads planning to a web-based Claude Code session. After running 10 Ultra Plans vs 10 local plans, he discovers that Ultra Plan secretly has 3 different planning modes assigned server-side, making it essentially an A/B/C test. The deep plan mode uses 4 sub-agents with a critique pass and is genuinely better for complex cross-app changes, but users have no control over which mode they receive. He recommends extracting the deep plan prompt and using it as a standalone custom skill instead.

Channel: Ray Amjad URL: https://www.youtube.com/watch?v=UNhA17l6CWw

🎯 Learning Objectives

By the end of this video, you will understand:

  • How the /ultraplan command works and what the web-based planning UI offers
  • The 3 hidden planning modes inside Ultra Plan (simple, visual, deep/multi-agent)
  • When Ultra Plan genuinely outperforms local planning vs when it’s equivalent
  • How Anthropic uses server-side A/B/C testing to evaluate planning prompts
  • How to extract and repurpose the deep plan system prompt as a custom skill

📋 Curriculum/Contents

  • ~0:00 — Introduction: what Ultra Plan is and why you might be underwhelmed
  • ~0:30 — Demo: running /ultraplan in Claude Code, getting session URL
  • ~1:00 — Web UI walkthrough: inline comments, thumbs up/down, approve plan
  • ~1:45 — Teleport options: run on cloud vs teleport back to terminal
  • ~2:30 — Benchmark results: 10 Ultra Plans vs 10 local plans, key differences
  • ~3:30 — Discovery: 3 hidden planning modes found via binary string analysis
    • Simple plan (same as local)
    • Visual plan (adds ASCII + Mermaid diagrams)
    • Deep plan / “free subagents with critique” (multi-agent exploration)
  • ~4:30 — Deep plan architecture: 4 sub-agents (code understanding, file identification, risk analysis, critique)
  • ~5:00 — A/B/C testing revelation: Anthropic controls assignment via remote config
  • ~5:45 — Hypothesis: Anthropic measuring plan acceptance rates to benchmark prompts and future models
  • ~6:15 — Use cases where Ultra Plan excels: blast radius analysis, dependency upgrades
  • ~6:45 — Recommendation: extract deep plan prompt as standalone custom skill (link in description)
  • ~7:20 — Summary: Ultra Plan is faster and better for multitasking, but mode assignment is opaque

📝 Notes & Key Takeaways

Main Insights

  • Ultra Plan is an A/B/C test in disguise — Anthropic assigns one of 3 planning modes server-side; users have zero visibility or control over which they receive
  • Deep plan = multi-agent + critique: 4 sub-agents handle code understanding, file discovery, risk identification, and a final critique pass — this is meaningfully better for complex tasks
  • Consistently 2x faster than local plan across all 10 test runs, likely due to cloud compute
  • Better for blast radius analysis: caught more issues when migrating tRPC v10→v11; outperformed local plan on cross-app changes
  • Not always better: for simpler isolated changes (e.g., swapping one model for another), Ultra Plan provided no advantage over local planning
  • Mermaid/ASCII diagrams appear when assigned to the visual plan variant — their presence/absence now has an explanation

Actionable Points

  1. Use /ultraplan for complex dependency upgrades or multi-file refactors where blast radius matters
  2. Leave inline comments on the web plan UI before approving — this refines the plan before execution
  3. Extract the deep plan system prompt from Claude Code binary strings and deploy it as a custom /deepplan skill (Ray shares this in description)
  4. When evaluating Ultra Plan output, consider you may have received the simple or visual variant — if the plan looks mid, it’s not necessarily Ultra Plan’s ceiling
  5. Use Ultra Plan for multitasking: fire off several plans simultaneously, review async in web UI

Personal Reflections

Add your own notes here after watching

⭐ Rating & Review

After completion:

  • Quality (1-5): _/5
  • Relevance (1-5): _/5
  • Would recommend: Yes / No
  • Best for: Claude Code power users who want to understand what’s happening under the hood with Ultra Plan, and practitioners looking to replicate the deep plan behavior as a custom skill

🏷️ Auto-Generated Tags

Content Analysis:

  • Type: video (YouTube content)
  • Topics: AI, Claude, tools, development, productivity
  • Complexity: Intermediate — assumes familiarity with Claude Code; some binary/reverse-engineering context is advanced
  • Priority: Medium — Ultra Plan is genuinely useful but still experimental; extracting the deep plan prompt is the most durable takeaway

Why These Tags:

  • AI + Claude — core subject is a Claude Code feature
  • tools — evaluates a specific tool feature and workflow
  • development — relevant to software development workflows using AI assistants
  • productivity — faster planning and multitasking angles
  • technical — includes binary string analysis and system prompt inspection
  • actionable — concrete recommendation: extract deep plan prompt as custom skill

Suggested Bases Filters:

  • Find similar content: type = video AND tags contains "Claude"
  • Find high-priority learning: priority = high AND status = inbox

🔗 Related Topics & Further Research

Related Searches:

  • Claude Code planning mode comparison (local vs Ultra)
  • Multi-agent critique patterns for software planning
  • Claude Code custom skills and system prompts
  • A/B testing in AI product development
  • tRPC v10 → v11 migration (referenced as test case)

Captured: 2026-04-08 Source: https://www.youtube.com/watch?v=UNhA17l6CWw Channel: Ray Amjad

Connection to Other Notes:

  • Related to Claude Code features and workflow optimization
  • The deep plan prompt extraction connects to custom skills development (kf-claude, kf-cli)