Context Is Architecture: Building Better Inputs for AI Systems

More context does not automatically produce a better answer. Long histories can bury the instruction that matters, repeat outdated facts, leak information across tasks, and spend tokens without improving the decision.

Good context is assembled, not accumulated.

Build a Typed Brief

Separate the material sent to a model by purpose:

1. Policy: stable rules the model must follow.
2. Task state: the current goal, completed steps, and open decisions.
3. Evidence: source passages or records relevant to this request.
4. Memory: user or workflow preferences that are valid in this scope.
5. Tools: available actions, parameters, and consequences.

Keep provenance with the evidence. The model should know whether a statement came from an authoritative record, a user message, a previous model summary, or an inference.

Select Before You Compress

Summarizing a noisy pile creates a smaller noisy pile. Filter by permissions, recency, relevance, and authority first. Then compress material that remains too large, while preserving identifiers and links needed for verification.

Ordering also changes behavior. Put non-negotiable policy where it cannot be mistaken for quoted content. Keep untrusted webpage text clearly separated from instructions. Present the immediate task and success condition close to the point of generation.

Evaluate the Context Builder

Test missing evidence, conflicting sources, stale memory, malicious document instructions, and tasks that should use no personal context at all. Measure not only answer quality but also citation accuracy and unnecessary disclosure.

The model may be the engine, but context is the road. A carefully built brief can make a modest model dependable; a chaotic one can make a powerful model confidently wrong.