At some point, adding more instructions makes the AI worse.

One thing you notice after enough hours building with AI: the longer the conversation gets, the more careful you have to be about the context.

That sounds backwards. A longer thread should mean a smarter assistant. It has the files, the decisions, the constraints, the corrections. In practice, past a point, the thread becomes a pile. The model is holding more, but it is not holding it well.

At some point, adding more instructions does not make the AI better. It can make the interaction worse.

Context is a budget, not a diary

Every message sits in a window with a hard size. The useful bits compete with the stale bits. A decision you reversed twenty turns ago is still in there, next to the rule you just added, next to a stack trace that no longer applies. The model does not get a clean briefing. It gets the whole tape.

Human conversations forget on purpose. AI conversations forget by overflowing, compressing, or quietly dropping the middle. You only see the failure when the output contradicts something you already settled.

More prompt is not more control

The instinct, when quality drops, is to add another paragraph of instructions. That is how you end up with a thread that contains three product briefs, two rejected approaches, and a style guide the model is no longer weighting. You feel like you were precise. You were loud.

The better move is often to start a new conversation with a short, current spec — or to treat the long thread as a draft of that spec, not as the working memory forever. Fresh context is not amnesia. It is a clean budget.

• When the task changes, start a new thread. Do not bolt a second project onto the first.
• Keep a running spec outside the chat. Paste the current version. Do not ask the model to remember the archaeology.
• If you find yourself restating the same constraint, the thread has already lost it.
• A shorter prompt against a known state beats a longer prompt against a muddy one.

You cannot manage what the window hides

Developers will keep using these tools more heavily. Longer sessions, more tools in the loop, more of the project stuffed into one chat. The failure mode will not be “the model is dumb.” It will be “nobody could see the context fill up until the work went sideways.”

That is one of the reasons we started working on TokenPulse — to make token and context usage visible instead of leaving people to guess what is happening behind the chat window. It is still early. The problem is not.

How long do you usually keep one AI conversation going before you start a new one?