For most of us now, some of the most important conversations happen inside ChatGPT, Claude, Gemini or Grok websites. (I wanted to talk about the privacy and everything, but some other day.)

Most of us treat these interfaces as chat. Even when we use them for serious work.

When we’re doing a long-running work, we seed the AI with some initial files: reports, data sets, images, background etc. And these documents make up the base for all our future conversations.

Initially, everything works well but as the time passes and conversations build up, the AI is overloaded with TMI (Too Much Information, courtesy: Michael Scott) And starts forgetting earlier details or outrightly makes up things.

What to do? How to make it more effective without any tricks. Because it is pure mathematics, no tricks work except for tricks of mathematics.

There are many things that we can do to make it more effective. But I want to share only two actionable things:

// 1. One idea per chat thread: Asking the AI about your child’s homework and adding your data to generate a report in the same chat is mathematically abusive to the LLM and your usage meter. Because a single thread carries all previous messages in that conversation.

If you’re sending your 10th message to the AI in the same chat thread, the previous 9 messages from you and the AI’s responses will be carried forward.

It is game-changer for most people because it saves the usage and mainly, INCREASES your AI’s attention to the concept being discussed.

// 2. Context distillation: This should be practiced when you’re working on long-running tasks. For example, you’ve got a 80 page report from where you need to check different things and use in your work. Context distillation significantly reduces your page count, mostly by 60%.

The idea is very simple: we ask the AI to remove all semantic words and condense the document. You don’t know how to do it, just use the prompt provided in the comment. Save the result as your source of truth and you’re done.