Ultrathink - Sequential Deep Thinking Mode

When this skill is invoked, use the Sequential Thinking MCP tool to methodically reason through the user's request. This produces higher-quality answers for complex problems by forcing structured, step-by-step analysis with the ability to revise and branch thinking.

How It Works

Take the user's prompt/question/task
Call the sequentialthinking MCP tool repeatedly to build a chain of thought
Each thought can build on, question, or revise previous thoughts
Synthesize the final answer from the completed thought chain

Execution Steps

Step 1: Identify the available tool

Check which sequential thinking tool is available:

mcp__sequential-thinking__sequentialthinking (primary)
mcp__thinking__sequentialthinking (alternative)

Use whichever is available. If neither is available, inform the user and provide the setup instructions (see "MCP Setup" section below).

Step 2: Estimate complexity and begin thinking

Based on the user's request, estimate how many thought steps are needed:

Simple questions: 3-5 thoughts
Moderate problems: 5-8 thoughts
Complex architecture/debugging: 8-15 thoughts

Start with your estimate but adjust freely as you go — add more thoughts if needed, revise earlier ones if they turn out to be wrong.

Step 3: Execute the thought chain

Call the sequential thinking tool iteratively. Each call should include:

thought: Your current reasoning step
nextThoughtNeeded: true (until you reach a satisfactory conclusion)
thoughtNumber: current step number
totalThoughts: your current estimate (adjust as needed)

Use these features when appropriate:

isRevision: true — when you realize a previous thought was wrong
branchFromThought — when you want to explore an alternative approach
needsMoreThoughts: true — when reaching the end but realizing more analysis is needed

Step 4: Synthesize and respond

After the thought chain completes (when nextThoughtNeeded is false), synthesize the results into a clear, actionable response for the user. Reference key insights from specific thought steps where helpful.

Guidelines

Think genuinely — each thought step should advance understanding, not just restate the problem
Be willing to revise — if thought #5 reveals that thought #2 was wrong, say so and correct course
Branch when there are genuinely different approaches worth exploring
The final answer should be better than what you'd produce without structured thinking — that's the whole point

MCP Setup

If the Sequential Thinking MCP tool is not available, ask the user to configure it.

For Claude Code CLI:

claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking

For manual configuration (add to MCP settings):

{
  "mcpServers": {
    "sequential-thinking": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-sequential-thinking"
      ]
    }
  }
}

Reference: https://github.com/modelcontextprotocol/servers/tree/main/src/sequentialthinking

ultrathink