Which levers you can pull and which you cannot. Your instructions, your files, project memory and the tools the model calls. When to start fresh, and why anything with numbers should go through code. Small, unglamorous habits that keep AI useful on long work.
Read the full chapter →Free to read on The Conversation Daemon, where every piece of the book is published.
Most of the friction people have with AI comes from trying to push on the layer they don't control while ignoring the layer they do.
The model's natural mode is approximation. Code's natural mode is exactness. When the question wants exactness, route it through code.
What the platforms keep, what you control, and what to build when that is not enough.
Why does your AI forget? Because it never remembered. Each message, the platform assembles a fresh stack of text inside a fixed budget, and what feels like memory is whatever made it into the stack. A plain walk through the layers: the budget, the librarian who fetches from your files, and the daily summary that drifts.
For the few who outgrow the platform's memory. A free database the AI can read and write over the web, holding a note on where you left off, the current state of the work and a running log. What it changes, what it is not, and a test prompt to see whether the door is open for you.
The same question through three AI models, and the same voice came back each time. The wording lives in the model, the character in your files, and the meeting only when you show up. Whatever you keep looking after is what survives the next model.