Karo Zieminski: Context engineering for product builders — the 2026 operating manual

Karo Zieminski published this guide in May 2026 to address a gap she observed in how product teams were approaching AI development: most organizations had invested in prompt engineering, but few had thought carefully about what the model actually knows when it answers — the deeper layer that determines whether any prompt will work reliably.

What the article is about

The article draws a clear distinction that matters for product decisions: prompt engineering determines what you ask the model, while context engineering determines what the model knows when it responds. Zieminski argues that these are not variations of the same skill but different layers of system design, and that most teams underinvest in the second because it requires more upfront architecture work.

The core mental model comes from a computing analogy attributed to AI researcher Andrej Karpathy and Shopify’s Tobi Lütke: the LLM functions as a CPU, the context window serves as RAM, and the person or system managing what gets loaded into that window acts as the operating system. The implication for product teams is that the context layer — what data gets surfaced to the model, in what form, and when — is as much a product design decision as the interface or the prompt itself.

Key insight on reliability

One of the article’s clearest arguments concerns the asymmetry between weak context and strong prompts. A well-crafted prompt will underperform if it operates on incomplete or poorly structured context. Conversely, solid context architecture can sustain reasonable results even with mediocre prompting. This asymmetry makes context engineering foundational rather than supplementary — a layer to design before optimizing prompts.

The LangChain framework and practical tools

Zieminski covers four canonical strategies for managing context effectively, which she attributes to the LangChain framework: Write (storing and updating relevant information), Select (choosing which stored information is relevant for a given query), Compress (reducing context to what fits efficiently in the window), and Isolate (keeping separate contexts from contaminating each other). Each strategy maps to different product scenarios — RAG pipelines, multi-session agents, and systems that personalize across users all call for different combinations.

The article also introduces Stanford’s ACE framework for enabling self-improving agents and provides concrete starting points: a simple file structure for an AI context library, hygiene practices for maintaining context quality over time, and architectural prompts for working through design decisions with a model.

Who it is useful for

Product managers and product designers building features that depend on LLMs will find this a practical orientation before getting into implementation details. The argument that context architecture is a product decision — not something to delegate entirely to engineering — will be useful for teams trying to define responsibilities around AI feature quality. The article is particularly relevant for anyone who has noticed that the same model produces wildly inconsistent output across similar inputs without understanding why.