The original title is "Context Engineering for AI Coding Agents: Practical Patterns for Reliable Output"

Short summary

The article argues that prompt engineering is outdated for AI coding agents — the real bottleneck is context engineering: what information is in the agent's context window and when. It offers concrete patterns including project-level instruction files (CLAUDE.md, .cursorrules) for persistent conventions, lightweight external memory directories for scar-tissue knowledge, and subagents to quarantine exploratory context from main sessions. The core insight is that agents fail not from bad phrasing but from missing or stale context.

• •Shift from prompt engineering to context engineering: agent reliability depends on what's in the context window, not phrasing
• •Use project instruction files for unchanging conventions and external memory dirs for hard-won scar-tissue knowledge
• •Delegate exploratory tasks to subagents to quarantine context and keep main sessions focused

Generated with AI, which can make mistakes.