The SuperClaude Field Guide

Claude Code is Anthropic's tool that lets Claude work directly on the files in a project folder: reading code, writing code and running commands. SuperClaude is a free add-on, made by an independent community, that gives Claude Code a set of ready-made working methods.

Without SuperClaude, you explain every time how you want Claude to approach a job. With it, you type a short command such as /sc:brainstorm and Claude follows a tried-and-tested procedure: it asks the right questions, involves the right expertise, uses the right tools and stops at sensible checkpoints so you stay in control.

The Five Building Blocks

Think of the framework as a stack. You touch the top layer; everything below supports it.

What Happens When You Type a Command

A Simple Way to Picture It

Imagine hiring a small consulting firm for each job. The command is the type of engagement: a design review, a build, a bug hunt. The specialists are the consultants who turn up. The helper tools are their reference books and instruments. The house rules are the firm's policies: check your work, never fake a passing test, never leave a job half done.

The Single Most Important Idea

Many commands stop on purpose. A planning command produces a plan and waits; it will not quietly start changing your code. That is a safety feature. Section 5, the safety ladder (page 17), shows exactly which commands look, which plan and which change files.

What SuperClaude Is Not

• It is not an Anthropic product, and Anthropic does not support it.
• It is not a separate app. It only works inside Claude Code.
• It is not magic. It is good working methods, written down. You still review what Claude produces.

Installation takes about five minutes. The core needs only Python; the helper tools are optional and can be added later, one at a time.

What You Need First

Install in Four Steps

terminalpipx install superclaude      # recommended
# or, if you don't have pipx:
pip install superclaude

terminalsuperclaude install           # 30 commands + 20 specialist files

terminalsuperclaude install --list    # shows every command and specialist
superclaude doctor            # health check
superclaude version           # should print 4.3.0

claude code/sc:help

Adding Helper Tools (Optional)

Helper tools make some commands much better: official documentation lookup, real browser testing, web research and memory between sessions. Every command still works without them. The superclaude mcp command installs them for you.

terminalsuperclaude mcp --list                          # see what is available and installed
superclaude mcp                                 # choose interactively
superclaude mcp --servers context7 --servers sequential-thinking
superclaude mcp --servers tavily --dry-run      # preview without installing
superclaude mcp --scope project                 # this project only (default: user)

Three tools need a free or paid key from their provider. The installer asks for it, or you can set it in your environment first:

The AIRIS Gateway (Advanced)

The installer also offers the AIRIS MCP Gateway, which bundles more than 60 tools, including Airis Agent and MindBase, behind one connection. It needs Docker running on your machine.

terminalsuperclaude mcp --servers airis-mcp-gateway

Updating

terminalpipx upgrade superclaude      # or: pip install --upgrade superclaude
superclaude update            # overwrite the old command files with the new ones

Optional: The Confidence-Check Skill

The /sc:agent controller can use a small "confidence check" skill that scores how sure Claude is before it starts building.

terminalsuperclaude install-skill confidence-check

Windows Notes

• Use pip or pipx in PowerShell. The install.sh script is for macOS and Linux (or Git Bash).
• If superclaude is "not recognized", run python -m superclaude install instead, or add Python's Scripts folder to your PATH.
• With several Python versions, pick one explicitly: py -3.12 -m pip install superclaude.

Uninstalling

terminalpipx uninstall superclaude                 # or: pip uninstall superclaude
rm -rf ~/.claude/commands/sc               # removes the 30 commands
claude mcp remove <tool-name>              # for each helper tool you added

This walk-through takes about twenty minutes and is completely safe: nothing changes your files until step 5, and even then you approve every change. Open Claude Code in any project folder, ideally one tracked by git.

claude code/sc:load

claude code/sc:recommend "I want to understand this project and then fix a small bug" --expertise beginner

claude code/sc:analyze . --focus architecture --depth quick
/sc:explain src --level basic

claude code/sc:troubleshoot "describe a small problem you know about" --type bug --trace

claude code/sc:troubleshoot "the same description" --type bug --fix
/sc:test

claude code/sc:git commit --smart-commit
/sc:save --checkpoint

A Routine Worth Keeping

There are 30 commands. You will use about eight of them most days. They are grouped below by what you are trying to do, and every card shows the same five things: what it does, a label for its effect on your files, the full syntax, a realistic example and the usual next step.

Every command sits on one rung of this ladder. The higher the rung on the page, the less it can touch. If you are nervous, stay on the first two rungs: they cannot change a single file.

Where Each Command Stops, and What to Run Next

Commands are designed as a relay: each one hands over at a defined point so you can review before the next leg begins.

A specialist (SuperClaude calls them agents) is a set of instructions that makes Claude think like a particular expert. Most of the time they switch on automatically, based on your words and the command you use. You can also call one by name.

claude code@agent-security-engineer "review the login flow for weaknesses"
/sc:implement "user profile API" --agent backend   # ask a command to favour one

Each specialist also has deliberate limits. A front-end architect will not design your database; it hands that to the back-end architect. If a specialist declines part of a job, that is why.

Panels of Named Experts

Two commands simulate panels of well-known authors. /sc:business-panel uses nine business thinkers; /sc:spec-panel uses software-engineering authors. Both choose suitable experts automatically, or you can name them with --experts. They are simulations of published ideas, not the real people. For specifications, the focus you choose picks the panel: requirements brings Wiegers, Adzic and Cockburn; architecture Fowler, Newman, Hohpe and Nygard; testing Crispin, Gregory and Adzic; compliance Wiegers, Nygard and Hightower.

Modes are working styles rather than jobs. They decide how Claude behaves while a command runs: asking questions, showing its reasoning, compressing its answers. They switch on by themselves when the situation calls for it, and you can force one with a flag.

MCP stands for Model Context Protocol, the standard way of plugging extra tools into Claude Code. SuperClaude knows how and when to use ten of them. None is required: every command falls back to Claude's built-in abilities when a tool is missing.

Choosing Between the Two Editing Tools

Tools That Work Well Together

How Research Rates Its Sources

Flags are small switches added to the end of a command, always starting with two dashes. They let you say "think harder", "use the browser", "be careful" or "keep it short" without writing a paragraph. Most also switch on automatically when they fit.

claude code/sc:analyze src --think-hard --c7          # deep analysis, with official docs
/sc:build --safe-mode --validate --focus security

Thinking Depth at a Glance

Behavior Modes

Change how Claude approaches the work.

Tool Switches

Turn specific helper tools on or off.

Thinking Depth

How hard Claude thinks before acting. Deeper costs more tokens.

Execution Control

How the work is carried out.

Output and Focus

What the answer covers and how it looks.

When Flags Disagree

Whichever command is running, Claude follows standing rules. Their spirit: evidence over assumptions, working code over documentation, efficiency over wordiness. They explain why it checks git first and refuses to switch off a failing test.

Never Compromised

Strong Preferences

Before Changing a File or Starting a Feature

Recipes are tested sequences of commands for real situations. Find the sentence below that sounds like you, then jump to the recipe.

I Want to…

How to Read a Recipe

Each recipe has an ID (such as F1), a level, a line saying when to use it, numbered steps, the result you can expect, and sometimes a tip. Run the steps in order. The gray line under each step tells you what it produces, so you know what to look at before moving on.

Find the symptom, read the likely cause, apply the fix. Most problems are solved by the first three rows plus a restart.

Agent

See Specialist. SuperClaude's word for an expert role.

API

A defined way for programs to talk to each other, for example a list of web addresses a server answers.

API key

A secret code a service gives you so it knows who is using it.

Architecture

The overall shape of a system: its main parts and how they connect.

Branch

A separate line of work in git, so changes don't affect the main version until you merge them.

Build

Turning source code into something that runs or can be shipped.

CI/CD

Continuous integration and delivery: automatic testing and deployment whenever code changes.

Claude Code

Anthropic's tool that lets Claude work directly on files in a project folder.

Commit

A saved snapshot of changes in git, with a message describing them.

Context window

How much text Claude can keep in mind at once in a conversation.

Core Web Vitals

Google's measures of page experience: loading (LCP), stability (CLS) and responsiveness (INP).

Coverage

The share of your code that tests actually exercise.

Dead code

Code that is never used and can be removed.

E2E test

End-to-end test: drives the real application the way a user would.

Flag

A switch starting with -- that changes how a command behaves.

Git

The most common tool for tracking changes to code over time.

Hop

One round of follow-up research, using what was found to search again.

MCP

Model Context Protocol, the standard for plugging extra tools into Claude Code.

Migration

A scripted change to a database's structure.

Mode

A working style, such as questioning or compressed, that changes how Claude behaves.

MVP

Minimum viable product: the smallest version that is still useful.

OWASP Top 10

The best-known list of the most serious web security risks.

PATH

The list of folders your terminal searches for programs.

pipx

A tool that installs Python command-line programs in their own tidy space.

Refactor

Restructuring code to make it simpler without changing what it does.

Repository (repo)

A project folder tracked by git.

Slash command

A shortcut typed in Claude Code that starts with /.

Specialist

A set of instructions that makes Claude act as a particular expert.

Specification (spec)

A written description of exactly what should be built.

Terminal

The text window where you type commands to your computer.

Token

A small chunk of text, roughly three-quarters of a word. Usage and limits are counted in tokens.

WCAG 2.1 AA

The accessibility standard most often required by law.

Print these pages and keep them beside your keyboard.

All 30 Commands

Every command name, syntax line, flag, installer option and version number in this edition was checked against the SuperClaude source code rather than against secondary documentation.

What Was Checked, and How

• The project repository SuperClaude-Org/SuperClaude_Framework, master branch, commit fe68862 (September 27, 2026): all 30 command files, 20 specialist files, 10 tool files, the flag, rule and mode files, and the installer's Python source.
• The published package on PyPI: latest version 4.3.0, requires Python 3.10+, MIT license.
• The changelog entries for 4.2.0 and 4.3.0, and the project's Windows and Serena troubleshooting pages.
• Every recipe uses only commands and flags present in that source. Where the source contradicts itself, both sides are shown.

Corrections to Revision 3.0

Known Gaps and Claims Not Independently Tested

• Token savings (for example "58K to 3K" for repository indexing, "30–50%" for compressed mode) are the project's own figures.
• Automatic triggers (such as modes switching on above 75% context) describe intended behavior; Claude may not always follow them exactly.
• SuperClaude changes often. Run superclaude install --list and superclaude version to confirm what you have.

textMade in Jurgistan. (2026). The SuperClaude Field Guide (Version 4.0) [Computer software]. https://Made-in-Jurgistan.github.io/superclaude-field-guide/

bibtex@software{Made_in_Jurgistan_The_SuperClaude_2026,
  author = {Made in Jurgistan},
  month = {10},
  title = {{The SuperClaude Field Guide}},
  url = {https://Made-in-Jurgistan.github.io/superclaude-field-guide/},
  version = {4.0},
  year = {2026}
}