sequential-thinking-skill

Sequential thinking for Claude Code — no MCP server required.
Branch. Revise. Adapt. Think deeper.

Quick Start · Features · How It Works · MCP vs sequential-thinking-skill

---

A Claude Code skill that replicates the Sequential Thinking MCP server with full feature parity. No MCP infrastructure needed — just copy a folder and go.

Claude gains structured, multi-step reasoning with branching, revision, and dynamic depth adjustment, all tracked by a lightweight TypeScript state machine that persists to disk.

Demo

Here's what structured thinking looks like in action:

💭 Thought 1/5
First, I need to clarify what type of auction we're dealing with.
I'll assume sealed-bid first-price unless told otherwise.
[1/5] history=1 next=true

🔄 Revision 3/5 (revising thought 1)
Wait — I assumed one format, but the question is generic.
I should cover multiple auction types. Changing approach.
[3/5] history=3 next=true

🌿 Branch 4/7 (from thought 2, ID: vickrey)
In a Vickrey auction, the dominant strategy is to bid your true
valuation regardless of the number of players. Exploring this path.
[4/7] history=4 branches=vickrey next=true

Each thought is a deliberate step — Claude can revise when assumptions are wrong, branch to explore alternatives, and extend when problems are deeper than expected.

Features

• Sequential numbered thoughts — structured chain-of-thought with formatted output
• Revisions — reconsider and correct previous thoughts without losing the original
• Branching — explore alternative paths from any thought, tracked by branch ID
• Dynamic depth — adjust totalThoughts up or down as the problem unfolds
• Hypothesis/verification loops — generate and verify solutions iteratively
• Persistent state — history survives across invocations via JSON on disk
• Full state inspection — dump complete thought history and branch details at any point

Quick Start

Prerequisites

• Claude Code CLI
• tsx — available via npx tsx (no global install needed)

Install

npx skills add https://github.com/thedotmack/sequential-thinking-skill --skill sequential-thinking

Or manually:

git clone https://github.com/thedotmack/sequential-thinking-skill.git
cp -r sequential-thinking-skill/sequential-thinking ~/.claude/skills/

That's it. The skill auto-activates when Claude detects problems that benefit from structured reasoning.

You can also trigger it explicitly with phrases like: "think through this step by step", "break this down", "use sequential thinking".

How It Works

The skill is powered by think.ts, a TypeScript state machine that:

1. Maintains an append-only thought history on disk (.think_state.json)
2. Tracks branches as named collections of thoughts forking from any point
3. Returns a compact status line to stdout after each thought ([3/7] history=3 next=true)
4. Renders formatted thought headers to stderr for visual feedback

Claude invokes it via tsx scripts/think.ts with CLI flags for each thought. The SKILL.md file instructs Claude on the protocol — when to branch, revise, extend, and terminate.

Commands

Command	Description
--reset	Clear state for a new thinking session
--thought "..." --thoughtNumber N --totalThoughts M --nextThoughtNeeded true/false	Submit a thought
--isRevision --revisesThought N	Mark as revision of thought N
--branchFromThought N --branchId "label"	Branch from thought N
--needsMoreThoughts	Signal that more thoughts are needed
--status	Dump full state (history + branches)

Output

Every thought prints a header and content to stderr, plus a compact status line to stdout:

💭 Thought 4/7
In a Vickrey auction, the dominant strategy is...
[4/7] history=4 branches=vickrey,english next=true

The --status command returns full JSON with fullHistory and branchDetails.

Comparison

How does sequential-thinking-skill compare to the official MCP server?

	Sequential Thinking MCP	sequential-thinking-skill
Setup	Configure MCP server in claude_desktop_config.json	Copy a folder to ~/.claude/skills/
Runtime	Requires running MCP server process	Standalone TypeScript script
Dependencies	@modelcontextprotocol/sdk, zod	tsx only
State persistence	In-memory (lost on restart)	JSON on disk (survives restarts)
Features	Thoughts, branches, revisions	Identical feature set
State inspection	Via MCP tool call	--status flag
Integration	MCP protocol	Claude Code skill protocol

Project Structure

sequential-thinking-skill/
├── sequential-thinking/
│   ├── SKILL.md                       # Skill definition and protocol
│   ├── scripts/
│   │   └── think.ts                   # TypeScript state machine
│   └── references/
│       └── example-session.md         # Worked example with all features
├── README.md
├── LICENSE
└── .gitignore

License

MIT