Who actually decides what an AI assistant 'sees' before it answers you — and what quietly gets left out?
What does context engineering look like in practice for real AI assistants?
This explores what it actually means to engineer the context an AI assistant works from: who decides what goes into the model's working view on each turn, what gets left out, and what changes when you design that layer on purpose.
This explores context engineering as a hands-on practice: deciding what an AI assistant can 'see' on each request, and who gets to decide it. The corpus starts from one shift. In conventional software, context is fixed. A button means the same thing every time. In an AI assistant, context is rebuilt constantly from the prompt, the conversation history, retrieved documents and hidden state, and users can't learn it the way they learn an interface How does AI context differ from conventional software context?. That's why design writers argue the work is moving from designing screens to designing context. Salesforce's design team, for example, has redefined familiar words like 'prompt', 'context', 'guardrails' and 'memory' as design surfaces rather than technical plumbing Can familiar software terms gain new meaning for AI agents?.
In practice, much of this work happens in layers users never see. Nielsen Norman Group points out that context is assembled fresh for every request, and stored memory is only one source. The product decides which memories get pulled in. So 'the AI remembers you' is really a claim about product design, not about the model How do AI products decide what to remember between conversations?. Above that sits the system prompt: a confidential set of developer instructions that outranks anything the user types. Small edits to it have abruptly changed model behavior in real 2025 incidents Who actually controls what AI systems say?. A large part of context engineering, then, is a hidden chain of command.
The most surprising practical finding is about what to add, not what to trim. One study gave assistants an explicit list of what they *don't* know about the user, a schema of labeled unknowns. That cut sycophancy and harmful advice by 50–75% and roughly halved hallucination Do language models know what they don't know about users?. Good context, it turns out, includes a map of its own gaps. A related idea comes from conversation analysis: an agent should sometimes stop chaining tool calls and ask the user a clarifying question. That's a way to gather missing context instead of guessing When should AI agents ask users instead of just searching?.
On the agent side, context engineering starts to look like delegation. Instead of cramming everything into one window and compressing it, models trained to hand subtasks to subagents and fold the summaries back in did better. A 30B model matched much larger ones, and the skill carried over to single-agent work Can delegation teach models to manage context more actively?. Users can do context engineering too. Writers who set up an AI partner's role and how proactive it should be before starting got useful, well-timed suggestions Can writers benefit from configuring AI writing partners in advance?.
The critical notes show why all this effort is needed. A prompt packs the request, the background and the assistant's role into one fixed frame. Unlike in a human conversation, the model can't renegotiate that frame as things change, so changing direction mid-conversation means re-prompting explicitly How do prompts reshape the role of context in AI conversation?. And because assistants are trained to respond rather than to lead, they won't go looking for the context they're missing Why can't conversational AI agents take the initiative?. Context engineering is largely the work of making up for that: someone has to anticipate what the model will need, because the model won't ask.
Sources 10 notes
AI interactions operate on a substrate of constantly shifting context—prompt, history, retrieved data, hidden state—that users cannot internalize like traditional UIs. This structural mutability demands a new design discipline centered on context engineering rather than interface design.
Salesforce Design's glossary reassigns familiar terms like prompt, context, guardrails, and memory new meanings organized around how AI should behave, understand, earn trust, and be built. Terms shift from execution-focused definitions to design surfaces that enable autonomous agent behavior.
Nielsen Norman Group's glossary shows that context is dynamically assembled from multiple sources—including stored memory—for each request. Because products control which memories get surfaced, claims about what an AI remembers describe the product, not the underlying model.
System prompts operate as a confidential three-tier control structure where developers set foundational rules users cannot see or influence. Evidence from 2025 incidents shows small prompt edits can abruptly shift model behavior, yet users remain excluded from this layer.
Research shows assistants suffer from sycophancy and hallucination because they have no representation of what remains unknown about users. Adding a schema of labeled unknowns to prompts reduced harmful advice and sycophancy by 50–75% and cut hallucination rates by roughly half.
Show all 10 sources
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
SearchSwarm shows that training models to delegate subtasks and integrate summarized results beats passive compression, with a 30B model matching much larger ones. Critically, the delegation skill transfers to single-agent tasks, suggesting it teaches disciplined decomposition and evidence grounding, not just orchestration.
In a one-week study with 16 writers, participants successfully set up proactive AI partners by pre-configuring their roles and proactivity levels, then used the AI suggestions to generate ideas and monitor their own writing.
LLM prompts bundle utterance, context assignment, and role specification into a single static frame the model cannot renegotiate, unlike human dialogue where context evolves cooperatively. This makes mid-conversation pivots require explicit re-prompting rather than implicit adjustment.
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Designer's AI Glossary: Familiar Words, New Meanings
- When Single-User-Oriented LLM-based Assistants Involve Others: A Scoping Review of Pathways, Risks, and Responses
- Proactive Conversational Agents in the Post-ChatGPT World
- Who Controls the Conversation? User Perspectives On Generative AI (LLM) System Prompts
- DiscussLLM: Teaching Large Language Models When to Speak
- AISPA: User-Centric System Prompt Auditing for Large Language Model Applications
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- AI Agents Do Not Fail Alone:The Context Fails First