Context engineering without the context dump
Better agent results come from selecting the right context at the right moment—not pasting an entire project into every prompt.
More context is not better context
A large context window can hold an impressive amount of material, but capacity is not the same as relevance. When an agent receives every brief, transcript, decision, and research note at once, the important constraint has to compete with a large amount of merely related text.
Context engineering is the discipline of deciding what the model should see for the task in front of it. The useful question is not how much can fit. It is which pages will change the quality of the next decision.
Retrieve context in layers
Start with metadata: titles, descriptions, folders, and status. That gives the agent a map without paying the cost of reading every page. Once it finds the likely source of truth, it can open the full content and follow links only when they are relevant.
This progressive approach keeps the working set small while preserving access to deeper history. It also makes retrieval explainable: you can see which page the agent selected and whether that choice was sensible.
- Search the relevant workspace before starting substantive work.
- Read the current project page before older research or archived notes.
- Pull supporting evidence only for decisions that depend on it.
Write descriptions for retrieval
A page title names the subject. Its description should explain the purpose, scope, audience, or outcome. That extra sentence is what lets an agent distinguish a launch decision log from a launch checklist without reading both documents in full.
Generic descriptions such as project notes or Markdown document waste the retrieval layer. Write the description for the future search that has not happened yet: what would someone need to know to choose this page from ten similar results?
Keep the current view current
Retrieval fails when the most discoverable page is stale. Update the living page when a decision changes, a blocker clears, or the next action moves. Version history can preserve what used to be true; the active page should describe what is true now.
A well-maintained workspace gives an agent a short path from search to action. That is the practical payoff of context engineering: less re-briefing, fewer contradictory answers, and more room in the prompt for the actual work.
