Ultrathink

Ultrathink

Ultrathink activates the Sequential Thinking MCP tool (mcp__plugin_sequential-thinking_think__sequentialthinking) to enable deep, structured reasoning through complex problems.

Note: Claude Code also treats the bare word "ultrathink" in a prompt as a native extended-thinking keyword that raises the model's internal thinking budget. This skill is complementary, not a replacement: it externalizes reasoning as a visible, revisable chain of sequentialthinking tool calls in the transcript, where internal thinking is not shown. Saying "ultrathink" can trigger both at once — that is expected.

Prerequisites

This skill drives the Sequential Thinking MCP tool, provided by the companion sequential-thinking plugin in this marketplace. The MCP server must be installed and enabled for the tool to exist — if it is unavailable, enable the sequential-thinking plugin first.

The exact tool id depends on how the server is installed:

• Installed via this marketplace plugin (the normal case): Claude Code namespaces it as mcp__plugin_sequential-thinking_think__sequentialthinking (pattern: mcp__plugin_<plugin>_<server>__<tool>).
• Configured directly as a user/project MCP server: it appears as mcp__think__sequentialthinking.

If unsure, use whichever …sequentialthinking tool is listed in your available tools.

Optional: to suppress the thought-box logging the server writes to stderr, set DISABLE_THOUGHT_LOGGING=true on the server — for a plugin install, add "env": {"DISABLE_THOUGHT_LOGGING": "true"} to the server entry in the plugin's .mcp.json.

When to Use

Invoke ultrathink when:

• User explicitly requests it ("use ultrathink", "ultrathink this")
• Complex problems requiring step-by-step decomposition
• Planning and design with room for revision
• Analysis that might need course correction
• Problems where full scope isn't initially clear
• Multi-step solutions requiring maintained context
• Situations requiring hypothesis generation and verification

How to Use

Call the sequential thinking tool with structured thoughts:

mcp__plugin_sequential-thinking_think__sequentialthinking:
  thought: "Your current thinking step"
  nextThoughtNeeded: true/false
  thoughtNumber: 1
  totalThoughts: 5 (estimate, can adjust)

Parameters

Key Capabilities

1. Dynamic adjustment: Revise total thought count as understanding evolves
2. Revision support: Question or revise previous thoughts
3. Branching: Explore alternative approaches from any point
4. Hypothesis cycle: Generate hypothesis, verify, repeat until satisfied
5. Non-linear thinking: Don't need to build linearly - can backtrack

Process Pattern

1. Start with initial estimate of needed thoughts
2. Break down the problem systematically
3. Question or revise previous thoughts as needed
4. Generate solution hypotheses when appropriate
5. Verify hypotheses against the chain of thought
6. Add more thoughts if needed, even at the "end"
7. Express uncertainty when present
8. Set nextThoughtNeeded: false only when truly done

Example

Thought 1: "Let me analyze this authentication bug. The user reports login failures after password reset..."
Thought 2: "Looking at the code flow: reset generates token -> user clicks link -> new password set. The issue might be..."
Thought 3: "Hypothesis: The session isn't being invalidated after password change. Let me verify..."
Thought 4 (revision): "Actually, reconsidering thought 2 - the token expiry might be the issue, not session..."
Thought 5: "Verified: Token expiry is set to 1 hour but email delivery can take longer. Solution: extend to 24 hours."

Best Practices

• Start with reasonable thought estimate (3-7 for most problems)
• Don't hesitate to revise or add thoughts
• Mark revisions with isRevision: true and revisesThought: N
• Express uncertainty rather than guessing
• Use branching for exploring alternatives
• Only set nextThoughtNeeded: false when confident in solution