Context Engineering: The AI Skill That Replaced Prompt Engineering

The five pillars of context engineering are selection, compression, ordering, isolation, and format optimisation.

Selection means choosing only the most relevant information to include. Not everything belongs in the context window. Compression means condensing long documents or conversation histories into concise summaries without losing critical details. Ordering means placing the most important information where the model pays the most attention, typically at the beginning and end of the context. Isolation means separating different types of information using clear labels, headers, or XML tags so the model can distinguish instructions from data. Format optimisation means structuring your input in the way the model processes most effectively, using bullet points for lists, tables for comparisons, and plain prose for narrative context.

Before loading everything you can think of into a prompt, ask yourself: what is the smallest set of information this AI needs to complete this task well?

This is called minimum viable context (MVC). It typically includes clear instructions and a stated goal, the specific user request, a few targeted examples showing expected behaviour, a minimal set of enabled tools, and the core facts tied to this particular request.

Start lean, test the output, and add context only when you identify a gap. Overloading the context window with irrelevant information actually degrades performance because the model spreads its attention across more tokens.

One of the most powerful context engineering techniques is just-in-time retrieval, where the AI fetches relevant information only when it needs it rather than having everything pre-loaded. Instead of pasting an entire 50-page document into the prompt, you provide a search tool and let the model retrieve only the sections it needs. Retrieval Augmented Generation (RAG) is the most common implementation of this technique, built into tools like ChatGPT with browsing, Claude with file search, and Gemini with Google Search grounding.

Context engineering also becomes essential when your AI needs to maintain coherence across multiple interactions. Short-term memory preserves conversation history within a single session. Long-term memory stores user preferences, past decisions, and learned patterns. For everyday users, this means using features like ChatGPT's memory, Claude's project knowledge, or Gemini's saved preferences.