Make Claude reason step by step before answering - catches what a direct reply misses.
Claude sometimes answers fast and confidently - and misses something important in the process. The default mode is: receive question, pattern-match, respond.
This setup adds the sequential-thinking MCP server to Claude Code. Instead of jumping to an answer, Claude works through the problem in explicit steps, commits to intermediate conclusions, and can backtrack if something doesn't add up. It surfaces contradictions and gaps that a direct answer skips over.
- Claude Code installed and running
- Node.js 18+ and npm (included with Claude Code)
- macOS (tested on macOS 14+)
open ~/.claude/mcp.jsonIf the file doesn't exist yet:
mkdir -p ~/.claude && touch ~/.claude/mcp.jsonPaste this into mcp.json. If you already have other servers configured, add the sequential-thinking block inside your existing mcpServers object.
{
"mcpServers": {
"sequential-thinking": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
}
}
}The package downloads automatically on first use.
The MCP config is only read at startup. In VS Code:
Cmd+Shift+P > Developer: Reload Window
Or quit and reopen Claude Code entirely.
Open a new chat and run:
think step by step: should I prioritize feature A or feature B if my goal is X?
If Claude responds with visible reasoning steps before the conclusion, it's working.
Claude activates sequential thinking automatically on complex requests. You can also trigger it explicitly:
think step by step: [question]
break this down step by step: [problem]
reason through this carefully: [decision]
Analyzing a job listing against your experience:
think step by step: here's the job description [paste text].
What requirements match my background, what's missing, what should I highlight?
Making a decision with multiple variables:
think step by step: I'm choosing between [option A] and [option B].
My priorities are [list]. Walk through which fits better and why.
Content strategy:
think step by step: my audience is [description].
Which angle would resonate more - [option A] or [option B]?
What would they already know, what would surprise them?
Debugging a plan or argument:
think step by step: here's my plan [describe].
What are the weak points? What did I miss?
| Works well | Limitations |
|---|---|
| Decisions with several competing variables | Slower than direct answers |
| Finding gaps in your own reasoning | Uses more tokens per request |
| Analyzing trade-offs | Overkill for simple factual questions |
| Complex prioritization (backlog, career, strategy) | Not needed for "what does X mean" |
| Catching contradictions in stated goals |
Asked Claude to help choose between two career directions using regular chat - got a pros/cons list. Asked the same question with sequential thinking - on step three, Claude identified a contradiction between two stated priorities that hadn't been noticed. It found that "I want stability" and "I want fast growth" were pulling in opposite directions, and asked which one actually mattered more.
The regular answer was correct. The step-by-step answer was useful.
Problem: sequential-thinking doesn't appear in Claude's available tools after reload
Fix: Validate your mcp.json is correct JSON:
cat ~/.claude/mcp.json | python3 -m json.toolProblem: Claude doesn't seem to be using step-by-step reasoning even with the server connected
Fix: Be explicit in your prompt. Add "think step by step" or "reason through this carefully" at the start. For very simple questions, Claude may correctly skip the full reasoning chain.
Problem: Responses are very slow
Fix: Expected behavior for complex problems. Sequential thinking works through multiple reasoning steps before responding. If speed matters more than depth for a given question, ask Claude directly without the step-by-step instruction.
Issues and PRs welcome - especially additional use case examples or prompt patterns that work well with this server.
MIT