Designing what the model sees
Context engineering is the practice of selecting the information an AI needs to do a job, and delivering it in a usable form.
Context here means more than the user’s request: the system prompt, earlier conversation, reference material, search results, tool descriptions, and notes accumulated during the work.
How it differs from prompt engineering
Prompt engineering is mostly about writing one instruction clearly. Context engineering is about the whole information environment surrounding that instruction.
The prompt “write an article” produces very different output depending on whether the model also has your previous articles, a reader profile, a style guide, a list of phrasings to avoid, and relevant sources. The work is less about finding better wording and more about assembling the material the model can reason from.
Is a large context window enough?
A model with a large context window can take in a lot, but holding information and using the right part of it are different things. Loosely related documents, stale instructions, and duplicated history can bury the constraint that actually mattered.
For agents that run for a long time, history and tool output pile up as work proceeds. Summarizing what matters, discarding what is stale, and preserving key decisions is part of context engineering too.
Putting it into practice
- State the goal and the definition of done up front
- Include only the material that is needed, with sources and priority made clear
- Separate standing rules from one-off requests
- Summarize decisions made along the way and carry them into the next step
- Be explicit about what the AI may access and what it may use
Keeping your own rules and past decisions in a system like Obsidian, then feeding the relevant parts to an AI when needed, is a practical example of the same idea.