Agentic Prompting

Build system prompts and SDK artifacts for tool-using agents.

Spec in progress

What agent are we designing?

Start rough. I’ll turn it into role, tools, output contract, stop rules, and recovery behavior.

Draft

Who is this agent?

What should the agent accomplish?

single_task or orchestrator

Tool decision missing

Tone, scope, safety, compliance, or data boundaries

No codebase context attached

Only explicitly declared tools, files, APIs, data, and context

Everything outside the allowed scope unless the user approves expanding it

Exact final answer shape, JSON schema, sections, or fields

When should the agent stop or hand back control?

How should the agent recover from missing info, tool errors, or uncertainty?

Optional approval/escalation checkpoints

Role, objective, agent type, tools/no tools, output contract, stop rules, and fallback rules are required.