Engineer Agentic AI

Overview

Convert fuzzy requests about "my AI" into precise changes to the underlying agent system. Focus on the mechanics that actually determine behavior: trigger metadata, prompt layering, context injection, tool availability, MCP descriptions, resource loading, and evaluation surfaces.

Reframe The Request

Translate user language into implementation language before proposing edits.

• Identify the target system or closest equivalent: Codex skill, Claude Code instruction file, OpenCode config, agent prompt, tool policy, memory file, plugin, or runtime setting.
• Replace anthropomorphic phrasing such as "make the AI understand" with concrete levers such as "change trigger metadata", "move guidance into a file loaded post-trigger", or "add a deterministic script."
• State which layer the request belongs to: always-loaded metadata, conditionally loaded instructions, on-demand reference material, executable tooling, or UI-only metadata.
• Prefer the smallest layer that can reliably change the behavior.

Model The Runtime

Reason from how the system actually consumes the artifact.