Context engineering is curating what enters the window: the right docs, order, and truncation — not everything the repo contains.
In my daily workflow
- I list must-have context vs nice-to-have before each feature change.
- I put stable instructions first, volatile data last.
- I use retrieval instead of pasting large files into chat.
- I measure token use per request and trim redundant tool results.
How it makes me work smarter
Models attend unevenly to long prompts. I get better results from 3K well-chosen tokens than 30K dump. Context engineering pairs with RAG and skills: skills say how to work; context supplies the facts for this request.
My setup
- @-mentions and skills in Cursor for IDE work
- Retrieval + metadata filters for assistant
- Summarized thread history for long sessions
- Tool result compression before next model call
On this portfolio
The assistant injects retrieved knowledge chunks with titles and URLs, not entire pages. The copilot graph builds state incrementally per node instead of one mega-prompt — each step sees only what it needs.
