Tell AI in three layers: who I am, what I'm working on

4,600 words of renovation notes, 16 living requirements, 3 files — you're about to send it all to the AI in one go.

Sunday afternoon, you sit at the folding table in the new apartment and open a fresh chat. Intro, the 80-square-meter old flat, the budget, the elder who feels the cold, the dog that slips on tile — plus six months of renovation diaries. You paste everything, then ask: “Hardwood or tile for the living-room floor?”

The logic feels clean: the AI answers vaguely when it lacks background, so you hand over everything you know. More detail, fewer misses. The more you give, the sharper the answer — that's how asking a person for help usually works, too.

It cites the elder, the dog, and the budget, line by line, and lands on hardwood. You put that option on the quote and put down a 3,000-yuan deposit the next day. That night, searching old notes, you find a line from four months ago:

The last place was also a ground-floor unit in the south. After the humid return-rain season, the hardwood cupped along the edges. Next time, check moisture protection first — then talk feel and price.

You trusted the answer because it really did quote several pieces of background. The problem: long-term facts, this renovation's constraints, and old lessons were crammed into one layer. The key fact in the old notes was a single line, sitting next to 16 current requirements.

The AI can split this pile for you first: keep long-term facts always on, update the current project on its own, and retrieve only the old notes that match this question. Layered external memory is there so the AI knows which kind — and which few items — to read this time. So what goes in each layer, and which one do you add first? Below we keep asking about the living-room floor, and watch the answer change layer by layer.

You can tell the three layers apart by how fast they change. The more stable it is, the earlier it sits. The more it grows, the more you retrieve it by question.

1. Layer 1 covers the you that doesn't change: role, current level, tools you use, how you like answers — into custom instructions. It stands alone because this stuff stays put for months, and every new chat needs it.
2. Layer 2 covers the project in front of you: goal, constraints, what you already tried, where you're stuck — into a project brief or the top of the project notes. It stands alone because a new project swaps this whole set.
3. Layer 3 covers experience that has piled up: each note records one choice, one result, and how to use it later, in a searchable library. It stands alone because notes keep growing. Pull only the few that match this question, and the key fact is easier to spot.

Adding layer by layer means: park the most stable facts first, then fill in the current project, then retrieve old notes by question.Each layer has its own update rhythm. When the renovation project changes, you don't rewrite the intro too.

Asking for all three layers at once slides you right back to that 4,600-word dump. Have the AI take one layer at a time. You confirm, then go on. The material is split from day one.

If layer 1 already makes the answer good enough, stop there. When the question starts depending on this project's conditions, add layer 2. When you need to check past choices and results, retrieve layer 3. Every time you add a layer, you should be able to say which fact it just brought in.

Keep notes wherever you already open every day, as long as you can find the relevant entries by topic before you ask. This is usage experience, not an experimental result. For notes that are easier to find later, see searchable notes.