Context Engineering
Curate the working set.
A clean prompt can still produce a bad answer.
A model works from the information available in its current interaction. That can include instructions, conversation history, files, examples, tool descriptions and results, retrieved material, and saved project state.
Context engineering is the practice of deciding what enters that working set, which source governs, and what must remain current.
Test context for
- Authority: Which source governs when materials conflict?
- Relevance: Does this information change the task or decision?
- Recency: Is its date, version, and status visible?
- State: Which decisions, changes, and open questions must survive?
- Boundary: Is the material appropriate, authorized, and safe to provide?
Purpose
Audience
Governing sources and versions
Approved decisions
Constraints
Current state
Open questions
Owner
Last updated
Ask for a context receipt.
Have the system identify the governing sources, conflicts, missing information, and prohibited actions. Correct that receipt before proceeding. Supplying a source does not guarantee that it was interpreted correctly.
People still resolve authority, privacy, and meaning.
Build one page of context.
Give it to a fresh interaction, request a context receipt, and compare that receipt with the source files.
The real skill is not giving the system everything. It is managing what should matter now.
Next: Workflow EngineeringSource check: Anthropic, Effective context engineering for AI agents.