# Ultrathink, Ultraplan, Ultracode, Ultrareview in Claude Code

> Claude Code's four ultra commands: Ultrathink reasoning, Ultraplan cloud drafting, Ultracode parallel agents, and verified Ultrareview bug detection.

Source: https://camnangai.com/en/for/developers/claude-code-ultra-commands  
Published: 2026-07-12

Ultrathink, Ultraplan, Ultracode, and Ultrareview are Claude Code's four "ultra" power commands —
  its highest-effort features, built for work a single normal response can't handle.

They turn Claude Code from a code editor into an agentic coding environment: architectural
deliberation, codebase-wide migrations, and verified bug detection.

What they share is a trade: more compute time and [tokens](https://camnangai.com/en/knowledge/what-are-ai-tokens) in
exchange for precision and autonomous coordination, by extending the reasoning budget, offloading to
Anthropic's cloud, or running many [sub-agents](https://camnangai.com/en/knowledge/what-is-an-ai-agent) in parallel.


  You don't reach for them on every task; you pull them out when the problem is hard enough to be
  worth the time and tokens.


New to the tool? Read [What is Claude Code?](https://camnangai.com/en/tools/what-is-claude-code) first, then come back.

![Cover illustration of Claude Code's four ultra power commands: Ultrathink, Ultraplan, Ultracode, and Ultrareview](https://camnangai.com/articles/bo-lenh-ultra-claude-code/01-infographic-en.webp)

## Ultrathink: telling Claude exactly how hard to think

**Ultrathink** gives you explicit control over how deeply the model reasons. Standard inference
predicts the next token immediately; extended thinking lets the model deliberate step by step before
answering, which stops it from pattern-matching its way through complex architectural logic. With
the maximum budget, Claude has room to run adversarial lines of thought and catch hidden bugs like
race conditions.

![Diagram of Claude Code's three thinking levels: think at 4,000 tokens, think hard at 10,000, ultrathink at 31,999](https://camnangai.com/articles/bo-lenh-ultra-claude-code/02-en.webp)

Drop these keywords anywhere in your prompt to set the reasoning depth:

- `think`: allocates 4,000 tokens for standard extended thinking, which is fine for moderate refactoring.
- `think hard`: allocates 10,000 tokens for deep thinking, best when weighing options and hunting
  edge cases.
- `ultrathink`: allocates the maximum 31,999 tokens, for root-cause analysis of intermittent
  failures or a major architectural shift.

While it runs, the model works through codebase structure, evaluates race conditions and timing, and
weighs competing implementation strategies rather than committing to the first one that fits. You pay for that in waiting, and in tokens, because a larger thinking budget is a larger bill.

```text
claude "ultrathink about migrating our auth module to microservices"
```

## Ultraplan: draft your plan in the cloud before you code

**Ultraplan** is a research-preview feature that hands a planning task from your local CLI to a
Claude Code web session. You draft complex changes in the cloud while your terminal stays free for
other work, and the browser gives you a richer review surface — inline comments and emoji reactions
on specific sections of the draft instead of replying to one long block.

![Ultraplan flow diagram: a local terminal hands a planning task to a cloud web session, you review it in the browser, then run it on the web or teleport it back](https://camnangai.com/articles/bo-lenh-ultra-claude-code/03-en.webp)

There are three ways to launch it from the CLI: run `/ultraplan` followed by your prompt, drop the
keyword `ultraplan` anywhere in a normal prompt, or pick the Ultraplan option when Claude finishes a
local plan and shows the approval dialog. The first two open a confirmation dialog before launching.

```text
/ultraplan migrate the auth service from sessions to JWTs
```

It requires Claude Code v2.1.91 or later, an authenticated Claude.ai account, and a project in a
GitHub repository. If Remote Control is active, it disconnects when Ultraplan starts. While the cloud
session works, your CLI shows a status indicator:

- `◇ ultraplan`: Claude is researching your codebase and drafting the plan.
- `◇ ultraplan needs your input`: a clarifying question is waiting in the web session.
- `◆ ultraplan ready`: the plan is ready to review in the browser.

Once the plan looks right, you either approve it to run in the cloud and open a pull request, or
"teleport" it back to your terminal. Teleporting gives you three options: `Implement here` (inject
the plan into the current conversation), `Start new session` (clear the context and start fresh with
the plan), or `Cancel` (save the plan to a file for later). Of those three I reach for `Start new session` most often, because a planning session's context is usually stuffed with research the implementation never needs.

## Ultracode: orchestrate many sub-agents in parallel

**Ultracode** (requires Claude Code v2.1.154 or later) combines extra-high `xhigh` reasoning with
automatic dynamic-workflow orchestration. Instead of working sequentially, Claude acts as an
orchestrator: it writes a JavaScript script that spawns many sub-agents, each handling an independent
piece of the job. The script lives in `.claude/workflows/`, so you can read, diff, or edit it before
it runs.

![Ultracode diagram: Claude orchestrates a script that fans work out to many parallel sub-agents, then reconciles the results](https://camnangai.com/articles/bo-lenh-ultra-claude-code/04-en.webp)

A few numbers show the scale:

- Up to 16 agents run in parallel, fewer on machines with limited CPU cores.
- A hard cap of 1,000 agents per run prevents runaway loops.
- A "Large workflow" warning fires when a run schedules more than 25 agents or is projected to pass
  1.5 million tokens.

It shines on large migrations — say, moving hundreds of files from REST to GraphQL. Claude splits
the files across sub-agents that work at the same time, then reconciles the results for project-wide
consistency. That 1.5-million-token threshold sounds distant, but when you fan a job out to a few dozen agents it arrives sooner than you expect. To kick off a session like that, set the effort level:

```text
/effort ultracode
```

## Ultrareview: catch bugs with multiple agents before you merge

**Ultrareview** (requires v2.1.86 or later) launches several independent agents in a remote cloud
sandbox. Unlike the local `/review`, which scans in a single pass, it makes each agent reproduce and
verify a finding before reporting it — which sharply cuts false positives.

![Ultrareview diagram: multiple agents in a cloud sandbox reproduce and verify each bug before reporting, versus a single-pass /review](https://camnangai.com/articles/bo-lenh-ultra-claude-code/05-en.webp)

You launch it from any git repository. With no arguments it reviews the diff between your current
branch and the default branch; pass a pull request number to review that PR instead:

```text
/code-review ultra
```

The older `/ultrareview` spelling still works as an alias for the same command.

On a real Flask API test, it pulled out exactly the classics people keep shipping:

- A hardcoded `SECRET_KEY`, opening the door to forged JWTs.
- Passwords stored in plaintext instead of hashed with bcrypt.
- SQL injection through direct string interpolation.
- The "silent None" bug: when an environment variable is unset, the JWT library still signs tokens
  with a `None` key and raises no error, a dangerous gap ordinary scans miss.

On cost: Pro and Max users get three free runs. After that, each run costs between $5 and $20
depending on the size of the diff, and Claude Code always shows an estimate you confirm before it
charges you.

## Which command should you use, and when?

You don't need an ultra command for everything, and overusing them just burns money. Don't call
`ultrathink` to look up loop syntax — you pay in tokens and waiting. For everyday work like
refactoring a basic function, `/review` and the `think` keyword are enough and cheap.

Reach for `/ultrareview` and `ultracode` when code is heading to production, especially changes that
touch authentication, the database, or payment flows. Spending $5–$20 on a multi-agent sandbox
review is cheap insurance against the cost of cleaning up a data leak in production.

## References

- [What Is the Ultra Code Mode in Claude Code? — MindStudio](https://www.mindstudio.ai/blog/what-is-ultra-code-mode-claude-code)
- [Claude Opus 4.8 Ultra Code Mode: Dynamic Workflows vs /goal — MindStudio](https://www.mindstudio.ai/blog/claude-opus-4-8-ultra-code-mode-dynamic-workflows-vs-goal)
- [Plan in the cloud with ultraplan — Claude Code Docs](https://code.claude.com/docs/en/ultraplan)
- [I Tried Claude Code Ultra Review — Joe Njenga (Medium)](https://medium.com/@joe.njenga/i-tried-claude-code-ultra-review-it-catches-bugs-before-you-even-merge-4d827ac667bd)
- [Orchestrate subagents at scale with dynamic workflows — Claude Code Docs](https://code.claude.com/docs/en/workflows)
- [Ultrathink: Tell Claude Exactly How Hard to Think — DEV Community](https://dev.to/rajeshroyal/ultrathink-tell-claude-exactly-how-hard-to-think-about-your-problem-i35)