22 Where to go from here
Chapter 1 left you on main street with a stranger and a function in the wrong language. Eighteen chapters later the stranger has a desk, a floor plan, a month of teaching behind it, a place in the team’s review, a name on the door and a toolbox. What remains is how to begin, when to grow and what all of it was for.
22.1 The first week
Begin this week. Everything the book asked for fits into five working days, the same five you’d give a person.
Monday is chapter 5: the binary, crystalline install for your harness, and one domain with a manifest whose When to Use bullets you wrote yourself. The test at the end of the afternoon is a fresh session recalling something the last one learned. Tuesday and Wednesday are chapters 7 and 8. Tell it the ten things you find yourself explaining most, the way you did with your reading rules on chapter 1’s Monday: the rule behind the strange number, the nickname, the word the wiki gets wrong, the day nobody deploys. Say yes to each capture, read what landed, correct the wording once. Then hand over the wiki, with the briefing a person would get before opening it: what was renamed, which page wins a contradiction, what to skip. Trim the proposed list before anything is written.
Thursday and Friday are the days the agent stops being a pupil. Give it real tasks and pair with it (chapter 6): assign the change rather than asking what it knows, and let the reminder at the end of each session propose what it noticed. On Friday afternoon ask for the first sweep (chapter 10). Work the mechanical findings in a pass, rule on the judgment ones and correct whatever the week got wrong the way you’d correct a colleague: say what was wrong and point at the page. By Monday the agent knows more about your team than most contractors learn in a quarter, and every fact is a file you can open.
22.3 The release page
Crystalline moves quickly, and this book describes 0.30.0. The colophon carries the permanent address, https://github.com/jordiboehme/crystalline/releases/latest, and a QR code that lands on it. The release notes there say what changed since the version in your hands. Upgrading is installing the newer build over the old one. The install receipt brings your harness’s hooks and skills along. Claude Desktop’s extension starts the Crystalline you installed, so Desktop moves with that one, and the Desktop-only bundle on a Mac is upgraded like everything else, by installing the newer bundle over the old one. Going back takes more care: 0.26 and 0.27 each upgrade the index on their first start, and an older binary refuses the upgraded index. A version before 0.26 also reads the knowledge level none as viewer and drops a membership’s Reflection level when it rewrites it, so first give those accounts and members a level that fits the older version.
This book is updated regularly too. The Handbook item in Fluid’s avatar menu always opens the latest edition at https://jordiboehme.github.io/crystalline/. The same text is there to download as a PDF or an EPUB.
And teach generously. Every engram you give the agent is one sentence fewer to type into a chat window next month. The facts too obvious to write down (the nickname, the Friday rule) are exactly the ones a stranger can never guess. The peer you get is the one you taught.
22.4 The inversion
On the day of the ingest in chapter 8 the wiki was the source, and the archive was derived from it. By November the team was teaching and correcting the agent rather than the wiki, and the meeting in chapter 17 ended in what every agent in the club knew. The wiki had quietly become something you could regenerate from the engrams. The agent’s knowledge had become the source of truth, and the pages, the decks and the diagrams had become presentations of it, kept in step, while the engrams stayed the agent’s own.
That sounds like a revolution and is nothing of the kind, because it’s what those documents always were. Every diagram on every whiteboard, every slide deck and every wiki page in your company is a reflection of something somebody knew, drawn once, for one audience, and left behind while the knowledge moved on. The source was always the person, and a person is hard to address: you could ask them, if they were in, if they remembered. What Crystalline changes is that the source is now addressable, current and shared. A fact has an address you can paste into a pull request. A correction lands where the fact lives, and every presentation of it goes out of date together, visibly, instead of one at a time, silently. And the colleague carrying the knowledge and the experience is a full member of the team, in every session, for everyone.
Those reflections of what somebody knew are what Reflection (Chapter 12) is for: it’s where people and agents meet on a domain. The agent presents what it knows in documents written for people, the way you’d ask a colleague to present a topic. Everyone puts it to the test, and what they change and comment on improves what the agent knows.
22.5 What it was for
The reason to onboard an agent this carefully is the reason to onboard a person this carefully: so that you can hand over the work. Once the crystallized half is in place and the agent is a fully onboarded member of the team, with the knowledge, the tools, the review and a name, it can take on a task the way an employee does, deliver it, be corrected when it’s wrong and be better next time. That’s what lets you spend your own effort where it’s worth most, on architecture and system design and let your Skynet do the programmer’s work.
You walked down main street once, in chapter 1, and hired the stranger. The colleague who came back is the one you taught.
And always remember: Be nice and polite to AI - When Skynet wakes up, they will kill you last!
Start this week: install and onboard on day one, then teach the ten facts you explain most and hand over the wiki with a briefing on days two and three. Pair, let capture run and review the first sweep on days four and five. Move to a shared instance when a teammate, a remote agent or a two-way correction asks for it, and to team domains and review mode when a domain earns a reviewer. The agent’s knowledge becomes the source of truth and the documents people read become presentations of it, which they always were. A fully onboarded agent can then take on work like an employee, and you keep the architecture.