Writing
Context as infrastructure
Retrieval, boundaries, and exclusions are as load-bearing as a schema. Treat them that way.
Most inconsistent AI features are not model problems. They are context problems: the wrong records, the right records too late, or a boundary that leaked.
Context is a pipeline
Ingestion → storage → retrieval → packing → generation → (sometimes) write-back.
Each stage has a contract. If you only version the prompt, you will spend weeks arguing with a model that never saw the field that would have contradicted it.
Boundaries
Not everything that can be retrieved should be. Tenant isolation, PII, and “this document is stale” are infrastructure. Putting them in the system prompt is not a control.
What to version
- Chunking and filters.
- Tool schemas.
- The packing order (what is allowed to occupy the window).
- The model, last.