Did Claude Code Just Get Plan Mode 2.0?
Did Claude Code Just Get Plan Mode 2.0?
Anthropic just shipped Ultra Plan, a cloud-based planning mode that was leaked last week and released for real two days later — and after a head-to-head test against standard plan mode, the verdict is "maybe." Ultra Plan is dramatically faster (30 seconds vs 5+ minutes), but it ignored the front-end-design skill I explicitly told it to use. That's a real problem. Here's the full comparison so you can decide whether to switch.
There's a ton of hype around Ultra Plan right now, but Anthropic's documentation is surprisingly thin — it basically says "plan mode, but in the cloud." My testing shows there's more going on under the hood, but the tradeoffs are real.
What Is Ultra Plan in Claude Code?
Ultra Plan is a cloud-hosted version of Claude Code's plan mode that runs in the browser instead of your terminal, is assumed (not confirmed) to use additional agents and architectural resources, and lets you edit plans with highlight-and-comment gestures in the browser UI. You invoke it with /ultraplan or by saying "ultra plan" inside an active Claude Code session.
The documentation only officially says three things: (1) it runs in the cloud, (2) you need a GitHub repo with at least one commit before you can use it, and (3) it's easier to revise the plan in the browser interface. Everything else — the multi-agent claim, the deeper planning resources — is inferred from testing, not stated.
That's a shame. If Anthropic is shipping something meaningfully more powerful than local plan mode, they should say so. Right now the docs undersell the feature.
How Do You Use Ultra Plan?
Run /ultraplan or type "ultra plan" in Claude Code on the latest version. You need a GitHub repo already created with at least one commit — a README alone counts. Claude Code handles the handoff: it asks if you're ready, launches a browser session, and gives you a link to open.
The flow:
- Invoke
/ultraplaninside Claude Code - Confirm the launch
- Follow the generated link to the browser
- Review the plan in the web UI
- Highlight any part of the plan to leave a comment or emoji
- Click "Approve Plan" to port it back to the terminal
Once approved, you have the option to execute within the current Claude Code session, start a fresh session with the plan preserved, or cancel. Starting a fresh session is the smart move for long tasks — you reset the context window but keep the plan.
How Fast Is Ultra Plan Compared to Local Plan Mode?
In a head-to-head test, Ultra Plan returned a full architecture plan in about 30 seconds. Local plan mode took over 5 minutes and needed a restart on the first attempt. That's a 10x speed difference on a mid-complexity project.
The test prompt: build a premium Kanban board web app from scratch. Use the front-end-design skill. Greenfield project with an empty GitHub repo. Include architecture, dependencies, and requirements.
- Ultra Plan: 30 seconds, ready to review in the browser
- Local plan mode: 5 minutes 30 seconds (second attempt — first attempt got stuck)
That speed gap alone is a real argument for Ultra Plan, especially on projects where you're iterating on planning and don't want to wait minutes per cycle.
Does Ultra Plan Actually Produce Better Plans?
On a head-to-head Kanban build, Ultra Plan's code quality was marginally better than local plan mode (a few hundred more lines and a slightly cleaner framework choice), but local plan mode produced a better final product because Ultra Plan ignored the front-end-design skill entirely. That's the biggest finding from the test.
I explicitly told both modes to use the front-end-design skill. Local plan mode used it. Ultra Plan did not, even though the prompt was identical.
The impact was visible in the final build:
- Local plan mode version had proper typography (Google Fonts), better card shading, small flourishes like an orange timer highlight, distinct color states for high/medium/low priority, and a more put-together visual hierarchy
- Ultra Plan version looked flatter — no Google Fonts, no custom design language, generic-looking card states, and noticeably less polish
Backend code quality was close. I spun up a third Claude Code session to compare the two codebases. It said Ultra Plan's architecture was slightly better. Gary Tan also picked Ultra Plan on architectural merits. But the visual gap is the user-facing gap, and that's where local plan mode won.
Why Did Ultra Plan Ignore the Front-End-Design Skill?
I don't know for sure, but in my testing, skills invocation in Ultra Plan has been unreliable across multiple projects, not just this one. If you're someone who relies on skills to shape outputs — and most experienced Claude Code users do — this is a dealbreaker until it's fixed.
The skill system is arguably Claude Code's most powerful native feature. Front-end design, skill creator, custom workflows — all of it depends on skills being invoked reliably. If Ultra Plan can't (or won't) hit that layer, you're trading speed for consistency, and consistency usually wins.
Is the Ultra Plan UI Better Than the Terminal?
Yes — the browser UI is genuinely nicer for iterating on plans because you can highlight specific sections and leave comments or emoji reactions directly on them. In the terminal, you have to describe what you want to change in natural language. In the browser, you point and comment.
That interaction model matters when your plan comes back with 20 steps and you want to tweak step 7 without rewriting the whole thing. You highlight it, leave a comment, and Claude picks up exactly what you meant.
The interface is probably half the value of Ultra Plan in its current state — especially for teams and for anyone who doesn't love text-based iteration on complex plans.
Should You Use Ultra Plan or Local Plan Mode?
Use local plan mode as your default when skills matter to your output, and try Ultra Plan on large, complex projects where the speed advantage and richer planning are worth the skill-invocation risk. This is a feature I'd test on your specific project type before committing.
Use Ultra Plan when:
- You're planning a large, complex project where more compute might find better architecture
- Speed matters more than skill-level consistency
- You want the browser UI for comment-based iteration
- You have a GitHub repo with at least one commit ready to go
Stick with local plan mode when:
- Skills are central to your output quality (front-end design, skill creator, custom workflows)
- You want absolute consistency with your existing workflows
- You're on a small-to-medium project where the extra planning power isn't needed
My honest take: Ultra Plan feels like it was pushed out slightly early to match the leak. The documentation is thin, skill invocation is inconsistent, and the Kanban test was probably too simple to showcase its upside. On a real complex project — the kind where you'd reach for GSD or Superpowers — Ultra Plan might genuinely separate from local plan mode. Test it there.
FAQ
Do you need a GitHub repo to use Ultra Plan?
Yes. Ultra Plan requires a GitHub repo with at least one commit — even a README counts. Without it, Claude Code won't launch the cloud planning session.
How fast is Ultra Plan compared to regular plan mode?
Ultra Plan returned a complete Kanban architecture plan in about 30 seconds. Regular plan mode took over 5 minutes and needed a restart on the first attempt. The speed gap is roughly 10x on mid-complexity projects.
Does Ultra Plan use more agents than regular plan mode?
Probably, but Anthropic hasn't officially confirmed it. The documentation only says Ultra Plan runs in the cloud and offers a browser-based review interface. Based on output quality and speed, it looks like there's more compute behind it, but the exact architecture isn't public.
What's the biggest problem with Ultra Plan right now?
Skills invocation is inconsistent. In my testing, Ultra Plan ignored the front-end-design skill even when the prompt explicitly required it — and this wasn't a one-off. If skills are central to your Claude Code workflow, this is a real blocker until it's fixed.
Is Ultra Plan worth switching to permanently?
Not yet. The speed and browser UI are real wins, but until skill invocation is reliable, local plan mode is the safer default. Keep testing Ultra Plan on larger, more complex projects where the extra planning power can actually show up. Expect rapid updates — this feature just shipped this week.
If you want to go deeper into Claude Code workflows and planning strategies, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.
