system-prompt-writer

System prompts guide agent behavior by defining role, context, and constraints. Effective prompts follow Anthropic's minimum effective information principle—providing just enough detail without over-specification—paired with structured templates using Markdown and XML to organize role, background, instructions, tool guidance, and output format for consistent, safe agent performance.

Tool guidance should include unambiguous decision logic and constraints directly in the system prompt. Use decision trees to clarify when tools apply, specify input requirements precisely, and document edge cases. Structure this section with clear precedence rules so agents consistently select and execute the correct tool without ambiguity.

System prompts containing code samples must use double-brace escaping `{{}}` instead of single braces `{}` to prevent template parsing errors. This critical escaping rule prevents the agent framework from misinterpreting code as template variables, which is a common source of KeyError failures in production workflows.

Multi-agent prompts require clear role separation and communication protocols. Define each agent's scope, decision authority, and handoff points explicitly. Use hybrid Markdown and XML structuring to distinguish coordinator logic from executor instructions, ensuring agents understand their position in planner-executor or similar architectures without task overlap.

Include only minimum effective information: the agent's role, relevant background facts, specific constraints, and decision criteria. Avoid over-context that confuses priorities. Apply the "right altitude" principle—pitch detail to the agent's actual task scope—and use structured templates to keep context organized and scannable without ambiguity.

Yes. System prompts enforce safety constraints directly by specifying prohibited actions, escalation rules, and boundary conditions. In customer service workflows, define what the agent can resolve autonomously versus when to escalate, which sensitive topics require human review, and what responses violate compliance—all within the system prompt structure.