Content
# Sequential Thinking Skill for Claude Code
[](LICENSE)
[](https://docs.anthropic.com/en/docs/claude-code)
[](https://agentskills.io)
**Token-efficient structured reasoning for complex analysis — replaces the sequential-thinking MCP tool with 73–95% fewer tokens.**
---
## The Problem
MCP tools inject their full schema (~950 tokens) into **every single message** in a conversation, whether the tool is used or not. Skills take a fundamentally different approach: only lightweight metadata (~50 tokens) is loaded per message, and the full skill content (~2,000 tokens) is loaded only when actually invoked.
In a typical 20-message conversation where sequential thinking is used twice:
- **MCP tool**: ~950 tokens x 20 messages = **~19,000 tokens**
- **Skill**: ~50 tokens x 20 messages + ~2,000 tokens x 2 invocations = **~5,000 tokens**
That is a **73% reduction** — and in sessions where the tool is never used, savings reach **95%**.
---
## Quick Install
### Standalone (recommended)
```bash
git clone https://github.com/elmaxid/sequential-thinking-skill.git
cp -r sequential-thinking-skill/skills/sequential-thinking ~/.claude/skills/
```
The skill is available immediately in your next Claude Code session.
### Plugin (future)
A `.claude-plugin/plugin.json` manifest is included in the repository for future marketplace support. Once the Claude Code plugin marketplace is available, installation will be a single command.
---
## What It Does
The skill provides a structured step-by-step reasoning framework. Each thinking step follows a consistent format:
```
## Step [N]: [Title]
- **Status**: CONTINUE | REVISE | BRANCH | CONCLUDE
- **Content**: The reasoning for this step
- **Next**: What follows
```
### Step Statuses
| Status | Meaning |
|--------|---------|
| `CONTINUE` | Proceed to the next logical step |
| `REVISE` | Re-examine a previous step with new information |
| `BRANCH` | Explore an alternative path in parallel |
| `CONCLUDE` | Final step — summarize findings and recommendation |
### Minimal Example
```
## Step 1: Identify the constraint
- Status: CONTINUE
- Content: The API rate limit is 100 req/s but peak load reaches 300 req/s.
- Next: Evaluate caching strategies
## Step 2: Evaluate caching strategies
- Status: BRANCH
- Content: Two viable paths — client-side cache (reduces calls by ~60%)
vs. server-side queue (smooths bursts but adds latency).
- Next: Compare trade-offs
## Step 3: Recommendation
- Status: CONCLUDE
- Content: Client-side cache with 30s TTL reduces peak to ~120 req/s.
Combine with a small server-side burst queue for the remaining 20% overflow.
```
---
## Token Comparison
| Metric | MCP Tool | Skill | Savings |
|--------|----------|-------|---------|
| Cost when idle (per message) | ~950 tokens | ~50 tokens | 95% |
| Cost per invocation | ~200 tokens | ~2,000 tokens | -900% |
| 20 messages, 2 uses | ~19,000 tokens | ~5,000 tokens | 73% |
| 20 messages, 0 uses | ~19,000 tokens | ~1,000 tokens | 95% |
> The MCP tool is cheaper per individual invocation, but its always-on schema cost dominates in real sessions. Skills pay only for what they use.
---
## Included Patterns
The skill ships with five reasoning patterns optimized for common engineering tasks:
| Pattern | Description |
|---------|-------------|
| **Architecture Decision** | Evaluate system design choices with explicit constraints, options, and trade-offs |
| **Root Cause Chain** | Trace a problem backward through its causal chain to the true origin |
| **Trade-Off Matrix** | Compare multiple options across weighted criteria to reach a justified choice |
| **Decision Tree** | Map conditional logic paths to handle branching scenarios systematically |
| **Incremental Refinement** | Iteratively improve a solution through successive rounds of critique and revision |
Each pattern provides a structured template that guides the reasoning process while remaining flexible enough for varied problem domains.
---
## Compatibility
- **Claude Code CLI** — fully supported as a native skill
- **Agent Skills open standard** — follows the specification at [agentskills.io](https://agentskills.io)
- **Operating systems** — works anywhere Claude Code runs (macOS, Linux, Windows via WSL)
---
## Contributing
Issues and pull requests are welcome. If you have a new reasoning pattern that would benefit the community, please open a PR with:
1. The pattern definition in the skill format
2. A brief example demonstrating its use
3. A short description for the patterns table above
---
## License
[MIT](LICENSE)
---
[Leer en Espanol](README.es.md)
