¿Qué es Context engineering? Definición y Contexto
Context engineering is the practice of choosing exactly what information goes into the model's context window — files, snippets, specs — so the AI has the right knowledge without noise.
Context engineering is the practice of choosing exactly what goes into the model's context window. Modern LLMs handle 200k–1M tokens, but more isn't better — irrelevant context dilutes attention, leaks across boundaries, and burns money. The engineering work is being deliberate about which files, snippets, examples, and rules the model actually needs.
In agentic coding, context engineering happens in three places: at session start (system prompt, AGENTS.md / CLAUDE.md, project rules), per-task (which files to read, which examples to load), and during the task (what to summarize and discard as the agent works).
The signal that context is wrong: hallucinated APIs, repeated mistakes the agent already saw fixed, or output that ignores conventions. The fix is almost always less context, not more — but more relevant.