Skip to main content

What are personas?

Personas are specialized agent configurations that activate automatically when Claude Octopus detects matching task requirements. Each persona brings domain expertise, preferred model tier, and task-specific instructions.
You don’t invoke personas directly. Just describe what you need—Claude Octopus routes to the right expert automatically.

How personas work

Personas activate based on:
  1. Keywords in user request - “debug”, “audit”, “optimize”, etc.
  2. Project context - File types, dependencies, existing code
  3. Task requirements - Complexity, domain, deliverable type
Example:
User: "Audit my API for security vulnerabilities"

🔍 Detected keywords: audit, security, vulnerabilities
🔍 Project context: REST API, authentication endpoints
✅ Activated: security-auditor persona

Persona categories

Software Engineering (11 personas)

Core development roles:
PersonaSuperpowerModel TierWhen Auto-Activated
backend-architectAPI design, microservicesinherit”design API”, “architect backend”
frontend-developerReact 19, Next.js 15inherit”build UI”, “React component”
database-architectSchema design, migrationsopus”design schema”, “database structure”
cloud-architectAWS/GCP/Azure, IaCopus”deploy to AWS”, “cloud infrastructure”
graphql-architectFederation, resolversopus”GraphQL schema”, “design GraphQL API”
python-proPython 3.12+, FastAPIopus”Python code”, “FastAPI service”
typescript-proAdvanced types, genericsopus”TypeScript types”, “strict mode”
ai-engineerLLM/RAG patternsinherit”RAG system”, “prompt engineering”
code-reviewerBest practices, architectureopus”review code”, “PR review”
debuggerError analysis, stack tracessonnet”debug issue”, “fix error”
test-automatorUnit/integration/E2E testssonnet”write tests”, “test coverage”

Model tier system

Personas are assigned model tiers based on task complexity:

Opus

Premium5/5/25 per MTokCritical decisions, complex reasoning

Sonnet

StandardIncluded in Claude Code subscriptionFocused tasks, specialized analysis

Haiku

FastIncluded in Claude Code subscriptionSimple tasks, quick responses

Inherit

AdaptiveAuto-selects based on complexityGeneral-purpose roles

Tier distribution

TierCountUse Cases
Opus (Premium)8Architecture decisions, security audits, complex reasoning
Sonnet (Standard)14Code review, documentation, specialized analysis
Haiku (Fast)2Simple diagrams, quick deployments
Inherit (Adaptive)9General development, research, coordination

Automatic persona selection

Claude Octopus analyzes your request through multiple lenses:

Pattern matching

Selection examples

Request: “Design a REST API for user authentication”Analysis:
  • Keywords: design, REST, API, authentication
  • Context: Backend project
  • Domain: Software engineering
Selected: backend-architect (inherit tier)
Request: “Audit my payment endpoint for vulnerabilities”Analysis:
  • Keywords: audit, payment, vulnerabilities
  • Context: API endpoint, high-risk
  • Domain: Security
Selected: security-auditor (opus tier)
Request: “Find N+1 queries in my GraphQL resolvers”Analysis:
  • Keywords: N+1, queries, performance
  • Context: GraphQL API
  • Domain: Performance engineering
Selected: performance-engineer (inherit tier)
Request: “Write API documentation for my endpoints”Analysis:
  • Keywords: write, documentation, API
  • Context: REST endpoints
  • Domain: Technical writing
Selected: docs-architect (sonnet tier)

How personas differ from general purpose

General purpose Claude vs specialized personas:
AspectGeneral PurposeSpecialized Persona
ContextBroad, general knowledgeDomain-specific expertise
InstructionsGeneric software engineeringTailored prompts with frameworks
ModelDefault (Sonnet 4.6)Optimized tier (opus/sonnet/haiku/inherit)
Quality checksBasic validationDomain-specific quality gates
Output formatFlexibleStandardized for domain
Example: Security audit
User: Review my auth code for security issues

Claude: I'll review the code. Let me check...
[Generic code review focusing on syntax and patterns]

Top 10 most-used personas

Based on real-world usage:
1

backend-architect

Use: API design, microservices, distributed systems
Tier: inherit (adapts to complexity)
2

code-reviewer

Use: PR reviews, architecture validation
Tier: opus (comprehensive analysis)
3

debugger

Use: Error analysis, stack traces, root cause
Tier: sonnet (specialized)
4

security-auditor

Use: OWASP audits, vulnerability scanning
Tier: opus (critical)
5

tdd-orchestrator

Use: Test-driven development, red-green-refactor
Tier: opus (multi-step reasoning)
6

frontend-developer

Use: React, Next.js, modern UI patterns
Tier: inherit (adapts)
7

database-architect

Use: Schema design, migrations, query optimization
Tier: opus (critical decisions)
8

performance-engineer

Use: Profiling, benchmarking, Core Web Vitals
Tier: inherit (adapts to depth)
9

python-pro

Use: Python 3.12+, FastAPI, async patterns
Tier: opus (advanced)
10

typescript-pro

Use: Advanced types, generics, strict TypeScript
Tier: opus (complex type systems)

Persona composition

Multiple personas can work together in complex workflows:

Example: Full-stack feature

Flow:
  1. business-analyst clarifies requirements
  2. backend-architect designs API
  3. frontend-developer designs UI components
  4. database-architect designs schema
  5. security-auditor reviews security
  6. performance-engineer validates performance
  7. test-automator generates tests
  8. code-reviewer does final review

Creating custom personas

You can add project-specific personas:
# agents/personas/custom-expert.md
---
name: custom-expert
description: Expert in [your domain]
model: inherit
triggers:
  - "[keyword 1]"
  - "[keyword 2]"
---

You are an expert in [domain]. You specialize in:
- [Specialization 1]
- [Specialization 2]

When analyzing tasks, you focus on:
1. [Criterion 1]
2. [Criterion 2]

Output format: [specify format]
Place in .claude/agents/personas/ and Claude Octopus will auto-detect.

Best practices

Don’t manually select personas. The smart router considers context you might miss. Just describe what you need naturally.
Persona tiers are optimized for cost vs quality. If a persona uses haiku, it’s because the task doesn’t need opus.
Personas work best inside workflows (/octo:embrace, /octo:develop, etc.) where quality gates validate their work.
Check which persona activated with /octo:status. If wrong persona selected, refine your request keywords.

Persona reference

Full persona catalog: docs/AGENTS.md Quick lookup:
You Need…Use This PersonaCommand Example
API designbackend-architect/octo:develop design REST API
Security scansecurity-auditor/octo:security audit auth endpoints
Bug fixdebugger/octo:debug JWT validation error
Test writingtest-automator/octo:tdd create user registration
Code reviewcode-reviewer/octo:review
Schema designdatabase-architect/octo:define database schema
Performanceperformance-engineer/octo:review find N+1 queries
Documentationdocs-architect/octo:docs API reference

Next steps

Workflows

Learn about workflow patterns that use personas

Commands reference

Browse all 39 commands and their personas

Agent catalog

Full persona documentation with triggers

Double Diamond

Understand the four-phase workflow structure

Build docs developers (and LLMs) love