AI starts every chat with amnesia and gets dumber the more clutter you pile into a single conversation, so the trick is to be deliberate about what context it sees. Good prompts give it a role, a task with context, and a clear output format, while the five context layers (custom instructions, memory, skills, tools, and your prompt) do the heavy lifting in the background. To stay out of the Dumb Zone, refresh the chat when the AI starts drifting, ask it to make a plan before tackling anything big, and use subagents to keep messy research out of your main conversation.
Sign in to view this.
Sign inSign in to view this.
Sign inAI wakes up with amnesia every time you open a new chat. It doesn't know you, doesn't know your agency, doesn't know your clients, your brand, or what you talked about yesterday. Every session you fill it with your instructions, the context about your company, relevant files and the poof, in the next chat it is all gone. It's basically a really capable new hire who shows up every morning having forgotten everything.
At the same time the AI seems to get dummer and dummer the longer you talk to it in a single session. How can we keep our AI smart without it forgetting all the context we have thought it before?
The easiest thing you can do to improve the results you get with AI is to improve the instructions that you give it as it is the easiest for us to control. A good prompt consists of three things:
Role: Tell the AI who it should act like, "A detail-oriented travel coordinator who specialises in luxury FIT trips", to put the AI in the right frame of mind. AI has gained through its training data a good understanding what FIT travel advisors do, it just needs to be nudged that it should be doing that type of work now.
Task+context: What is the exact task that we are doing right now and what context / documents will be required to perform that task well? Even stating the exact steps you would take to perform the task nudges the AI in the right direction
Besides you instructions AI agents accumulate context and information into the conversation. These are the five layers where that context comes from.
Each layer adds context. The more of them you set up, the more the AI knows about what it needs to do and how it should do the task. But it can also clutter the AI's focus.
AI has finite context called the context window usually 200,000 words/token to 1,000,000 words/tokens. AI can only take in information as long as there is space in the context window for that information. At the same time, as the context window fills up the AI starts to feel dumber until we reach what AI engineers call, the dumb zone.
This is when there is so much clutter in the AI brain that it can't make up what is actually important any more. The longer your conversation runs, the more stuff piles up. Old instructions, corrections from five messages ago, context from a different task you started with. The AI starts blending things together and drifting from what you actually want.
"Don't think about a pink elephant."
What's in your head right now? A pink elephant. I works the same way. Everything you put into a conversation stays there, and the AI can't really ignore it even when it stops being relevant.
Because the context window is finite and the AI gets dummer and dummer the more (irrelevant) information is in there, we need to carefully pick what the AI knows. Using tools, like searching the internet, creating documents or reading your emails quickly fill up the context window and are expensive opperations that get us to the dumb zone fast. The following tips and tricks are there to get the right results with enough context without going into the dumb zone:
AI starts every conversation with amnesia. Context arrives from five places: custom instructions, memory, skills, tools, and your prompt. Without enough of it you're stuck with generic outputs or misunderstandings, but with too much context you will end up in the Dumb Zone, with the AI drifting away from what you actually want. Get deliberate about what's in the context window. Set up the layers that never change, give richer context in the moment using voice mode, start fresh more often than feels natural, create a plan before doing anything big or important and use a subagent for large preliminary research tasks.
Output format: "Two short paragraphs. Match the style of this example:". "A PDF document that includes all images from the folder x". Don't allow the AI to fill in the blank of what the exact output should be. Do you want a 3 word slogan instead of a sentence, you should say that.
