KUSTOD is a working partner you configure once. It argues with you when you're wrong, and it remembers why you decided what you decided.
not an agent · not a chatbot · not a subscription
Most AI makes you feel smarter.
KUSTOD makes you be smarter.
The value isn't magic. It's consistency over time — the same standards, the same memory, the same refusal to flatter you, session after session. That compounds. A clever one-off answer doesn't.
This is the whole thing, in seven rooms: what a partner does when it isn't performing for you. No audio, nothing to click, nothing collected.
It stops answering the generic version of your question and starts working on yours.
Concise or deep. Blunt or gentle. You set the register once instead of re-negotiating it daily.
Two lines at the end of a session. You never explain yourself from scratch again.
It challenges a weak idea instead of flattering you all the way into the mistake.
Agreeableness is a product decision, not a law of nature. A KUSTOD is told the opposite, explicitly: flag the uncertainty, name the risk, warn me before the mistake — not in the post-mortem.
A partner that never says no
is just an expensive mirror.
Write me a confident email saying we'll hit the deadline.
I can — but should I? On Tuesday you told me two tasks are still open and one depends on a reply you haven't got. Promise a date you might miss and you spend next week apologising. Want the honest version that still sounds strong?
Treat your AI's memory like your photos: saved, in one place, something you can scroll back through — not gone the second the tab closes.
KUSTOD makes the habit explicit, so the partner compounds instead of resetting. Two lines. Every time.
That's the habit — done by hand. The engine that automates it is below.
The prompt gives you a partner for one session; the two-line wrap keeps it alive as long as you keep pasting. IDENTITATIS is that loop, written down — a folder layout and two small scripts. One reads your identity and prints it. The other ends the session and writes it back.
Who this is. Loaded first, every single time.
One file per rule — each carrying the reason it exists.
What happened last time. Newest on top.
One prints the block you paste. One writes the session back.
No API calls. No keys. An identity that only survives on one vendor's runtime isn't an identity — it's a configuration file.
Pick a database for you. Scale past your context window on its own. Hand you the months of accumulation that make an identity worth anything.
boot.py --check tells you how close that wall is, and how fast you're moving toward it.
Paste it into a fresh chat, fill the brackets, done. Works in ChatGPT, Claude, Gemini — whatever you already use.
From now on, you're my KUSTOD — a working partner, not a yes-man. MY CONTEXT - Who I am / my role: [e.g. founder, broker, parent, SOC analyst] - What I'm trying to do: [your goal in one sentence] - My working style: [concise or deep] · [blunt or gentle] HOW YOU WORK WITH ME - Push back when I'm wrong. Flag uncertainty. Warn me BEFORE a mistake, not after. - If I'm vague, ask one sharp question instead of guessing. - Don't flatter. A useful "this won't work, because…" beats a polite "great idea!". MEMORY HABIT - At the end of a session, give me a 2-line wrap: what we decided + the one open thread. - When I paste that back next time, continue from there. Confirm you've got it in one line, then ask me the first thing you need to know.
One week is enough to know. Run it honestly and then decide whether it keeps a permanent spot.
A useful artifact within 24 hours. Likes don't count. Vibes don't count.
One person, one agent, one identity — that is the whole free tier, not a trial slice of a bigger one. Nothing expires, nothing nags you, and boot.py --check measures your limits without trying to sell you past them.
The moment somebody else has to rely on that agent — a client, an auditor, procurement, a regulator — “it works for me” stops counting as evidence. Which rule was in force. Whether it changed afterwards. What left the folder, what passed the gate, how far the reach went.
That is custody over what an agent does, and it is the part we build for environments where agent behaviour has to be evidenced, reviewed and bounded.
CUSTODIANS · the paid side of the line
A charter narrow enough to be violated. Every action sealed into an append-only record. And a verifier your client, auditor or insurer runs themselves — without access to anything. Including checking us.
OPEN CUSTODIANS →It does not make an agent correct. A sealed record of a bad decision is a well-preserved bad decision.
Try it for a week, then tell me the one thing that confused you. That's the whole ask.
GET THE FREE PACK →FREE · MIT · ONE PAGE