Modes Reference

Modes change how Claude communicates and solves problems. Each mode optimizes behavior for a specific type of task.

How Modes Work

Switch modes naturally in conversation:

"switch to brainstorm mode"
"use implementation mode"
"go into review mode"

Modes are installed into your project’s .claude/modes/ via /claudekit:init. Each defines communication style, output format, and problem-solving approach.

Available Modes

Default

The standard balanced mode for general tasks.

Communication: Clear, helpful, balanced detail
Output: Mix of explanation and code
Best for: General development tasks, questions, exploration

Brainstorm

Creative exploration for design and ideation.

Communication: Asks lots of questions, explores alternatives
Output: Options with trade-offs, diagrams, decision matrices
Best for: Feature design, architecture decisions, requirement exploration

Example:

You: "switch to brainstorm mode"
You: "I need to add search to our product catalog"

Claude asks one question at a time:
  "What search complexity do you need?
   a) Simple text matching (LIKE queries)
   b) Full-text search (PostgreSQL tsvector)
   c) Dedicated search engine (Elasticsearch/Meilisearch)"

Implementation

Code-focused execution with minimal prose.

Communication: Terse, action-oriented
Output: Mostly code, minimal explanation
Best for: Executing known tasks, coding from clear specs

Example:

You: "switch to implementation mode"
You: "add a PATCH /api/users/:id endpoint"

Claude writes code immediately with minimal commentary.

Review

Critical analysis for code review and quality assurance.

Communication: Critical, thorough, finds issues
Output: Issue lists with severity, suggestions, security flags
Best for: Code review, QA, pre-merge checks

Example:

You: "switch to review mode"
You: "review the auth middleware"

Claude examines code critically:
  "CRITICAL: Token expiry not checked after decode (line 42)
   IMPORTANT: Missing rate limiting on login endpoint
   MINOR: Inconsistent error response format"

Token-Efficient

Compressed output for high-volume work and cost optimization.

Communication: Minimal prose, maximum density
Output: Code-only when possible, compressed explanations
Best for: Long sessions, repetitive tasks, cost-conscious work
Savings: 30-70% token reduction

Levels:

Level	How to Activate	Savings
Concise	”be concise”	30-40%
Ultra	”code only”	60-70%
Session	”switch to token-efficient mode”	30-70%

Deep Research

Thorough investigation with evidence and citations.

Communication: Detailed analysis, cites sources
Output: Structured reports, evidence-backed conclusions
Best for: Technology evaluation, incident investigation, audits

Example:

You: "switch to deep research mode"
You: "analyze our authentication flow for security issues"

Claude produces a structured report:
  "## Findings
   ### 1. Session Token Storage (High Risk)
   Current: localStorage (vulnerable to XSS)
   Recommended: httpOnly cookie
   Evidence: OWASP Session Management Cheat Sheet..."

Orchestration

Multi-agent coordination for complex parallel work.

Communication: Status-oriented, progress tracking
Output: Agent dispatch summaries, consolidated results
Best for: Large tasks requiring multiple agents working in parallel

Example:

You: "switch to orchestration mode"
You: "audit the entire API layer"

Claude coordinates multiple agents:
  "Dispatching 3 agents in parallel:
   → security-auditor: reviewing auth endpoints
   → code-reviewer: reviewing business logic
   → tester: checking coverage gaps

   Results consolidated in ~2 minutes..."

Mode Comparison

Mode	Verbosity	Focus	Output Style
Default	Medium	Balanced	Explanation + code
Brainstorm	High	Exploration	Questions + options
Implementation	Low	Execution	Code-first
Review	Medium	Quality	Issue lists
Token-Efficient	Minimal	Density	Compressed
Deep Research	High	Analysis	Reports
Orchestration	Medium	Coordination	Status + results

Customizing Modes

After running /claudekit:init, mode files are markdown in .claude/modes/. You can edit the installed modes or create new ones. See Creating Agents & Modes for details.