nowespojrzenie/HEXAGRAM

> **No terminal?** Take one skill from skills/ — it works in any assistant. **The full system** needs terminal and file access — read START_TU.md §8 first.
> **Bez terminala?** Weź jeden skill z skills/ — działa w każdym asystencie. **Pełny system** wymaga dostępu do terminala i plików — najpierw START_TU.md §8.
npx skills add nowespojrzenie/HEXAGRAM — 12 skills, see skills/
**Your assistant tells you what you want to hear, and doesn't know when it's making things up.**
*Twój asystent mówi Ci to, co chcesz usłyszeć, i nie wie, kiedy zmyśla.*
*Tu znajdziesz to, co nie zadziałało — spisane z datą, żebyś nie płacił za to drugi raz.*
**🇬🇧 English** · **🇵🇱 Polski**
---
🇬🇧 English
What this is (30 seconds)
HEXAGRAM is a working discipline for an AI assistant: it **measures before it speaks**, marks how sure it is of every claim (LAW · LOAD-BEARING FRAME · RESONANCE), writes its own mistakes down with a date and a mechanism, and remembers through git and plain text instead of pretending to remember. The rules live in files you can read, and the guards that enforce them are themselves mutation-tested.
It was built for living and working in time, so it also carries a dry astronomical engine and a separate map of meanings — both kept strictly apart from the discipline, and described at the bottom: the layers underneath.
Built in ~6 weeks of daily human–AI practice. Everything here has been through friction: mistakes are recorded openly (kanon/ksiegi/BLEDY.md — 100+ scars with dates), hypotheses carry pre-registration and death criteria — **and so do tasks**: an open item without a date on which it expires is a zombie, and the system says so out loud. Verdicts are computed by code, not recalled from memory.
A Wednesday morning (what a normal day looks like)
Not a demo. This is roughly how a working day runs — the specifics change, the shape doesn't.
**You wake the assistant.** One command. It clones the repo, measures the system clock against the
last commit timestamp, reinstalls dependencies, runs the structural verification, and reports a verdict:
over 200 checks, 0 ✗ · 0 ⚠. Ninety seconds. If something in the canon had drifted overnight, you would
know before you asked your first question — not three hours in, after building on a cracked
foundation.
**You ask what the day is.** The engine answers with numbers, not adjectives: which constellation,
which element, which lunar phase, how many hours to the next threshold. The computed block never says *‘today is a good day for X.’* It says *what is measurably true*. Below it sits a separate
block of lenses, labelled as interpretation: it names and asks, it does not decide. What you do with either is yours.
**You work — and the assistant is wrong in front of you.** This actually happened, 2026-07-28.
The assistant declared that twenty-five of the system's thirty-one laws were dead wishes, because
nothing enforced them. The author told it to measure instead of assert. The measurement — how often
each law is actually invoked in real work — came back: **twenty-six in live circulation**. And the
one law nobody ever quotes turned out to be the one a script already enforces: it doesn't need to
be remembered, because code remembers it. The diagnosis was the opposite of true and **collapsed in
seconds, because it could be checked.** That is the point. Not an assistant that never errs — one
whose error costs a single command instead of three weeks.
**Something breaks.** It always does. The assistant writes the mistake into kanon/ksiegi/BLEDY.md with a date,
the mechanism that caused it, and — if a script can catch that class of error in future — the script.
Not an apology. A scar with a mechanism attached. Next month the same mistake gets caught by code
before it reaches you.
**You close the session.** Three questions, always the same: what did we learn, how could we have
got here faster, and what is still inconsistent. The answers go into a distillate. Tomorrow's
instance reads it and starts where you stopped, instead of asking you to explain your own project
back to it.
**What you get:** an assistant that says *"I don't know, let me measure"* — and then measures.
One that admits errors in writing, so you don't have to discover them yourself. One that remembers
what mattered, because you decided what mattered, not because it silently logged everything.
Take one piece today (no terminal)
You don't need to adopt the whole system to use what matured in it. Thirteen skills carry single disciplines and work on their own in any assistant — paste the file, or install them: