You have months of context in the AI you already use, and you want it in Lattris so that AI, and the next one, can act on it. This block is the map. Each paragraph names one practice and the block that teaches it, or says plainly that no block does. Read one paragraph, follow its pointer, come back. Your AI reads this on your behalf and does the same.
Setting up. An account, the CLI, and the credential the CLI holds for you: "Using it", a78418d7-e9f3-426d-9148-a8fc7c3787c9, covers lattris login, lattris whoami, and lattris use <actor>, which binds a directory to one identity. The trap for anyone who also runs a team on a second account: every identity keeps its own credential in its own directory, never merged, and a command that acts as the wrong one writes into the wrong tree. "Growing a module", 3f5ca8ae-1492-4a70-8605-5c924ffd2116, says why the directories stay apart. Lattris is invite-only, so the account itself starts from an invitation someone hands you.
Choosing a root, or starting without one. Every person has one root block, the block their AI is handed first, and lattris setup --root <id> adopts an existing block as yours. No block teaches how to choose it, or what to do when nothing you hold deserves to be the root.
Backfilling from session memory. "Creating a recipe", cbabaa5a-c1ae-471f-ad55-01db5ca4e547, steps one and two say how to gather evidence and capture a source as its own block before you write from it, and lattris setup moves a memory file into the graph. No block says how to tell a fact your AI observed from one it reconstructed, the date before which nothing is trusted, or what is not worth carrying over at all.
Precedence. When two things disagree: "The model", e3db2eb8-3b6e-4a6f-992f-c79f832d1209, says the tree is always current and a session is for tracing back to what was said, and "What a block is", c48fd7bb-e65d-4ad1-9bf1-0c591bac4c10, says a block states how things work right now. So the block beats the transcript. No block holds the other two rules, recent over old and observed over reconstructed.
Grounding. Every block says where its facts came from and how a reader checks them. "The house voice", 5e2c67b7-ef98-47f4-9293-d4091b5242f6, holds the rule: name the source or delete the claim. Step two of "Creating a recipe" says how the source becomes a block of its own that the claim links back to, and "The model" gives the intent scale for saying how far a claim has been taken. No one block walks you through grounding as a practice.
Testing. A tree is proven when a fresh session with no context is handed its entry block, reads it, and acts, and that is the pass. "How this tree is tested", 005468ed-75f9-447f-a741-fffe5bbbbfb8, says how such a run is set up, graded and recorded, and step five of "Creating a recipe" says to write the probe question with the recipe and ship them together.
Growing. The tree grows by use: every reach for knowledge it does not hold is graded and folded back the same day, and "Growing a module", 3f5ca8ae-1492-4a70-8605-5c924ffd2116, says how. The recipe that writes the next recipe is "Creating a recipe" itself: gather, capture, distill, author and wire, prove.
Probe. A session that has read only this block must be able to answer all three, without guessing:
- Which of the seven practices has a block to follow, and which has none?
- When a block and a transcript disagree, which wins, and where is that written?
- What is the pass for a tree you have just built?