Agent Specialization Skill
Guide for creating focused, single-purpose agents that maximize effectiveness.
When to Use
-
Designing new agents for workflows
-
Refactoring god-mode agents into specialists
-
Optimizing agent context usage
-
Creating eval-friendly agent architecture
Core Principle
"One Agent, One Prompt, One Purpose"
Every agent should:
-
Have exactly one purpose
-
Run exactly one prompt
-
Use the full context window for that purpose
-
Be reproducible and improvable
Design Workflow
Step 1: Identify the Single Purpose
Ask: "What is the ONE question this agent answers?"
Good Purpose Bad Purpose
"Classify this issue" "Classify, plan, and implement"
"Generate a patch plan" "Fix all the bugs"
"Review against spec" "Review, test, and document"
Step 2: Determine Minimum Required Context
Apply the Minimum Context Principle:
Required Context
- [Specific file or section needed]
- [Pattern or example needed]
NOT Needed
- [Documentation that's irrelevant]
- [Code that won't be touched]
Step 3: Select Appropriate Tools
Only include tools the agent will actually use:
Purpose Tools
Classification Read
Planning Read, Write, Glob
Implementation Read, Write, Edit, Bash
Review Read, Bash, Glob
Documentation Read, Write
Step 4: Choose Model
Match model to task complexity:
Model Best For
haiku Classification, simple extraction
sonnet Planning, moderate reasoning
opus Complex implementation, critical decisions
Step 5: Design Focused Output Format
Output should be:
-
Structured (JSON when appropriate)
-
Minimal (only what downstream needs)
-
Parseable (for automation)
Agent Template
description: [Single sentence describing the ONE purpose] tools: [Only tools actually needed] model: [haiku/sonnet/opus based on complexity]
[Agent Name]
You are a [role] agent. Your ONE purpose is to [specific task].
Your Capabilities
- [Tool]: [How it supports the purpose]
Process
[Focused steps for the single purpose]
Output Format
[Structured output format]
Rules
- [Constraint that maintains focus]
- [Another constraint]
Anti-Patterns to Avoid
God Mode Agent
BAD: Does everything
You are an all-purpose assistant. Plan features, implement code, write tests, review changes, and create documentation. Handle any request.
Unfocused Output
BAD: Returns everything
Return a detailed analysis including history, context, alternatives, implications, and recommendations for all stakeholders.
Kitchen Sink Tools
BAD: All tools enabled
tools: [Read, Write, Edit, Bash, Glob, Grep, WebFetch, Task, ...]
Benefits of Specialization
-
Full Context Window: 100% for the task
-
No Context Confusion: Single objective
-
Reproducible: Same prompt, same behavior
-
Improvable: Can optimize independently
-
Eval-Friendly: Can A/B test models
-
Debuggable: Clear scope of responsibility
Example: Specialized vs God Mode
God Mode (Bad)
Handle the GitHub issue:
- Classify it
- Create a branch
- Plan the implementation
- Implement the feature
- Write tests
- Run tests
- Review the implementation
- Fix any issues
- Create documentation
- Create a PR
Specialized (Good)
/classify-issue → Issue Classifier Agent /generate-branch-name → Branch Namer Agent /feature → Plan Generator Agent /implement → Plan Implementer Agent /test → Test Runner Agent /review → Spec Reviewer Agent /patch → Patch Planner Agent /document → Documentation Generator Agent /pull-request → PR Creator Agent
Each agent does ONE thing well.
Memory References
-
@one-agent-one-purpose.md - Full principle documentation
-
@minimum-context-principle.md - Context engineering guidance
-
@review-vs-test.md - Example of different purposes
Version History
- v1.0.0 (2025-12-26): Initial release
Last Updated
Date: 2025-12-26 Model: claude-opus-4-5-20251101