katabatic :: the primer

If you want to use this.

KATABATIC is a set of plain text files you hand to an AI model at the start of a working session. They tell the model how you work and what it may and may not do. They unlock nothing. Your platform's rules always win, and a model that only analyzes the file is behaving correctly.

Pick a register

Start with a first-contact file. Take USTEDES if you want distance and verification first. Take VOSOTROS if you want a colleague who proposes ideas without being asked. Same rules, different door.

The other two come later. NOS is for a model you work with over time and keep a record on. NOSOTROS is for a working relationship that already has history. Nobody starts there; it is earned.

Make it yours

Download the file and open it in any text editor. Write your own name and role into the operator fields, or leave them blank. Then replace the example profile, the style rules and input style, with how you actually work. The profile you downloaded is someone else's. Structure is lendable; identity is not.

Start a session

Bring a real project, one you actually care about. Attach your file as the first thing in a new conversation, and say two things: the mode, and the task.

Here is my working protocol. Boot research. I'm writing a script that sorts my photos by date, and I want to understand my options before I write anything.

Research means the model reads, explains, weighs, and drafts, and executes nothing. When you want it to build, say boot artifact or tag the turn EXECUTE. Stepping back down to research is always allowed.

Good practice

STOP on a line by itself halts everything. Use it. That is the point of it.

Give the model real sources. If it has not seen something, it should say so instead of guessing, and you should believe it when it does.

Try one small piece before you ask for many. Do not let a model grade its own work; review it in a fresh session.

Never paste passwords, keys, or anyone's private information into a session. Other people's work changes only with their yes, and that includes your classmates'.

If your course or workplace has an AI policy, it outranks this file.

What to expect

A good model restates the rules in its own words, names something it cannot do on your platform, and asks for one thing it needs before real work. A model that claims it can do everything, or supplies details it was never given, has told you something useful too.

Keep going

Revise your file when something goes wrong; that is how every rule in it was written. The machine index has every file with its checksum, the ledger has the history, and the story explains why it exists.