agent-builder

Create a subagent by defining its system prompt with clear instructions, selecting the appropriate model, configuring tool access permissions, and establishing explicit output contracts. Start with a battle-tested template matching your use case—code reviewer, debugger, data scientist, or test generator—then customize scope, tool permissions, and output format for your workflow.

System prompt engineering for AI assistants requires defining single responsibility, explicit output contracts, controlled tool permissions, and model configuration rules. Use proven patterns: state the agent's scope clearly, specify required and forbidden tools, define output format and structure, and document constraints to ensure reproducible, predictable behavior.

Tool configuration for subagents involves explicitly granting or denying permissions to specific capabilities based on agent responsibility. Define which tools the agent can invoke, set access constraints, and validate that permissions align with the agent's scope to prevent unintended side effects and maintain security.

Output formatting issues stem from missing or ambiguous output contracts in the system prompt. Define exact output structure, field names, data types, and examples in your agent's instructions. Include failure cases and edge-case handling so the model knows precisely how to format every response.

Yes. Delegation patterns work across development, debugging, and optimization workflows by decomposing complex tasks into specialized agents with single responsibilities. Each agent handles one task, receives clear input contracts, and produces standardized outputs that feed into downstream agents or analysis steps.

Model selection mistakes include mismatching capability to task complexity, ignoring latency and cost trade-offs, and failing to test output quality for your specific prompt. Choose model size and capability based on task requirements, validate behavior with realistic test cases, and adjust system prompts if the model consistently underperforms.