agent-builder

Design a Claude subagent by defining its single responsibility, scope boundaries, and measurable outcomes. Write a detailed system prompt specifying the agent's role, processes, and output format, then configure minimal necessary tool access and select the appropriate Claude model (Opus, Sonnet, or Haiku) based on task complexity.

A subagent system prompt must define the agent's role, its core responsibilities, step-by-step processes it follows, expected output formats, and quality standards. Include specific examples of inputs and desired outputs, constraints on behavior, and any domain-specific terminology or requirements the agent needs to understand.

Select Claude model based on task complexity and latency requirements. Use Haiku for simple, fast tasks with minimal reasoning; Sonnet for balanced performance on moderately complex workflows; Opus for reasoning-intensive, multi-step problems requiring deep analysis and reliability.

Grant only the minimum tools necessary for the subagent to complete its task. Common patterns include Read for file access, Grep for searching content, and Bash for git operations. Define tool permissions narrowly—for example, restrict Bash to git commands only—to maintain security and prevent unintended actions.

Yes, subagents can delegate tasks to other specialized subagents through defined delegation patterns. Structure your architecture so each subagent owns a single responsibility, with clear handoff points and standardized input/output formats between agents to ensure modular, reliable multi-agent workflows.

Inconsistent outputs typically stem from vague system prompts, unclear output format specifications, or mismatched model selection. Tighten your system prompt with explicit examples of desired output, define a rigid output schema, and verify the model you chose matches your task's reasoning requirements.