# Context Engineering

> The design and continual selection of the information an AI model receives so it can act well within a limited context window.

- HTML version: https://robbiepalmer.me/ideas/context-engineering
- Source: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents

Context engineering manages the full input available to a model at inference time. That includes
system instructions, tool definitions, retrieved documents, examples, message history, tool
results, and working notes. Prompt wording sits alongside all of these other inputs.

More context is not automatically better. Irrelevant logs, overlapping tools, stale documentation,
and repeated instructions compete with the evidence needed for the current decision. The useful
target is the smallest set of information that lets the model act correctly, with a way to retrieve
more when the task demands it.

Agents make the problem dynamic. Each tool call creates new state, long tasks outgrow one context
window, and the value of earlier details changes as work progresses. Search, compact summaries,
structured notes, and clear file or graph links let an agent recover detail without carrying the
entire history in every turn.

## Questions it prompts

* Which information could change the next decision?
* What can be retrieved when needed instead of included up front?
* Which source wins when instructions or facts conflict?
* Does evaluation test retrieval and context assembly as well as the final prompt?

## Related ideas

- [Feature Engineering](https://robbiepalmer.me/ideas/feature-engineering.md): The work of turning raw observations into model inputs that represent the signal needed for a prediction.

## Where it appears

- Technology: [Basic Memory](https://robbiepalmer.me/technologies/basic-memory.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt