Ultrathink

Ultrathink is the per prompt version of max effort. Drop the word into any single message and Claude immediately shifts into its heaviest single pass reasoning mode for that request only. It exists so you can get deep thinking on one specific question without parking the entire session on a slow setting. The rest of your work stays fast while the hard problem gets the attention it needs.

Unlike the /effort command that holds across messages, ultrathink is surgical. You keep medium as your default and only invoke ultrathink when a particular task earns the extra cycles. The model thinks harder, checks its logic, and surfaces tradeoffs inside that one response. Then it drops back to your normal level for the next ask.

It is not the same as ultracode. Ultracode spins up multiple agents that review each other. Ultrathink remains one model doing one long think. It is also not a replacement for good task framing. Shouting ultrathink at a vague prompt still produces vague but slower output. The word is a booster, not a miracle.

Designers regularly mix it up with the persistent dial. They think typing ultrathink once sets the level forever. It does not. The boost dies after that response. This confusion leads to people overusing the word and then wondering why their session still feels slow on routine tasks.

A product designer at Vercel in 2025 used ultrathink when deciding between two approaches for their new streaming UI components. The normal prompt gave a quick opinion. Adding ultrathink forced the model to map latency implications, perceived performance patterns, skeleton screen timing, and progressive disclosure rules before committing. The response included a clear recommendation with exact token budget numbers and fallback states. The rest of the session stayed on medium so the team could keep iterating on visuals at full speed.

Another case came from a solo design engineer auditing a large Figma file. He needed to understand why certain variant states kept breaking across breakpoints. A normal prompt summarized the surface symptoms. Ultrathink made the model trace the four layer file structure, map every conditional rendering path, and identify the exact token layer causing the drift. The output read like a senior engineer who had spent an hour in the file with him.

Use ultrathink when one specific step in your process needs deeper scrutiny but the surrounding work does not. Drop it into prompts about architecture decisions, complex bug diagnosis, or high stakes copy that must carry exact brand voice. Avoid it on bulk operations, simple research synthesis, or any task where speed matters more than exhaustive analysis. The extra latency on trivial asks just frustrates you.

The tradeoff lives in precision. Used correctly it gives you occasional senior level thinking without slowing your entire workflow. Used as a crutch it becomes expensive punctuation that adds cost with diminishing returns. The skill is reading the task and knowing when one hard think will actually move the project forward.

Teams that use it well often build a short list of trigger conditions. Wide solution space, irreversible decision, client facing output. Anything outside that list stays on normal effort. That filter turns ultrathink from a novelty into a scalpel.

Ultrathink is the difference between asking an intern for an opinion and asking a principal designer to sit with the problem for thirty minutes. Know when to invoke it.