Chapter 2 · Lesson 3
Manage Context Over a Long Task
Keep extended AI work grounded when a conversation grows, changes direction, or moves between tools.
After a long conversation, the AI starts forgetting early decisions and reintroducing ideas you already rejected.
Conversation history is not perfect memory
Long chats can contain competing instructions, outdated drafts, and more material than the model can weigh reliably. Even when the full conversation fits, important constraints may receive less attention as new messages accumulate.
Create a working brief
Periodically summarise the current goal, accepted decisions, rejected directions, open questions, and next task. Ask the model to confirm the brief, then use it as the clean starting point for the next stage or a new conversation.
Keep the source of truth outside the chat
Store approved facts, final decisions, and current files in a document or system your team controls. Chats are useful workspaces, but they should not become the only record of why a decision was made.
Build the mental model
Long work stays reliable when decisions are compressed into a maintained brief.
Separate state from conversation
The chat records everything that was said; the working brief records what is currently true. Keep goals, accepted decisions, constraints, and open questions in the brief.
Refresh at meaningful boundaries
Create a new brief after a major decision, before changing tools, or when the task moves from research to drafting. Do not wait for obvious drift.
A launch plan has grown across a long chat with several rejected directions.
Remember what we discussed and write the final launch plan.
Create a working brief with: current goal, approved audience, accepted messages, rejected directions with reasons, confirmed evidence, unresolved questions, and next deliverable. Quote the relevant decision for any uncertain item. I will correct the brief before you draft the plan.
The brief turns hidden conversational state into an artifact the user can inspect, correct, save, and move to another tool.
What belongs in the source of truth
Store approved facts and decisions outside the chat in the system your team already trusts. Link to source files rather than relying on copied fragments when possible.
Date the brief and identify its owner. When facts change, replace the old state instead of appending another conflicting instruction to the conversation.
Create a checkpoint brief for active work
Pause one ongoing AI task and write a compact state document before continuing.
Goal: [...] Accepted decisions: [...] Rejected: [...] Evidence: [...] Open questions: [...] Next task: [...] Updated/owner: [...]multiple choice
What is the best response when a long AI conversation starts drifting?
Choose the answer that best fits the principle from this lesson.
Use a maintained brief as the source of truth for long AI-assisted work.