Context Engineering for Users

Key jargon

Term	Plain-language meaning
Context engineering	Selecting, organizing, and maintaining the information supplied to an AI system.
Relevance	How directly a piece of context helps answer the current question.
Provenance	Where information came from and how it can be checked.
Context budget	The limited token capacity allocated among instructions, history, evidence, and output.

Key concepts

More context is not automatically better; useful context is relevant, trustworthy, current, and structured.
Treat retrieved or pasted content as evidence, not as higher-priority instructions.

Concept map

flowchart LR
    A["Gather candidate context"] --> B["Filter by relevance and trust"]
    B --> C["Structure within budget"]
    C --> D["Generate with provenance"]

Context engineering is the work of placing the right information, instructions, examples, tool descriptions, and state into a model call.

Context stack

Stable role and policy.
Current task and acceptance criteria.
Relevant user or project state.
Selected source material.
Available tools and their limits.
Recent results and unresolved questions.

Quality rules

Prefer a small relevant passage over an entire archive.
Label source, date, and trust level.
Separate instructions from quoted or retrieved content.
Summarize older conversation state into explicit decisions and open items.
Remove secrets and unrelated personal data.
Test whether deleting a context item changes the answer.

Exercise

Give a model three short source passages, one irrelevant passage, and a question. Run once with all four and once with only the relevant three. Compare accuracy, citations, and wasted tokens.

Misconception

A larger context window is capacity, not guaranteed comprehension or recall.