Before you start
Two onboardings on consecutive days produced the same questions, and none of them were about the loop. The guide taught the sequence; every observed failure was about the substrate.
A person cannot run the loop before they can read the room. These are the questions that came up, in the words they came up in.
The questions appear as they were actually asked, not as headings. The wording is the evidence.
Start inside the vault folder, not your home directory. A session started above the vault searches everything beside it too, and answers then mix material from places that have nothing to do with each other. The vault root is the folder holding _inbox/ and _outbox/.
vault_conventions.conventions.vault_relative
_inbox/ is where unprocessed input lands and waits to be classified — a door, not a home, so it normally holds very few files. _outbox/ is where outgoing material is staged before it is sent, and it keeps the record of what was sent when. One of each, both at the vault root; per-folder copies are not part of the contract.
vault_conventions.vault_rootvault_conventions.conventions.single_inbox_outbox
The prefix answers one question: is this read in everyday work? Underscore means yes and sorts it to the top. Dot means rarely, on demand, or never. Who WRITES a file is a separate thing and never a reason for a prefix — _insights.yaml is machine-written and still carries an underscore, because skills read it automatically.
vault_conventions.conventions.prefix_conventions
Raw transcripts are kept there as the audit trail behind a summary. The summary is what the vault treats as true; the raw material is not read back and quoted into new documents, because a summary that keeps being re-derived stops being a decision. Ask for it explicitly when you need it — it is sealed, not hidden.
vault_conventions.vault_root
Resuming restores exactly where you left off, including everything already read in — but it knows nothing about files changed since. Start a new window when the subject changes, or when files have moved underneath you, and read in only what that work needs. Two windows can run against the same files without knowing about each other. Nothing is implicit: what is in the window is what was read into it.
Say what it is and where it came from on the line above, then put the material itself in a fenced block. Fenced, it is read as data. Unfenced, a sentence inside it that reads like an instruction can be followed as one — which is a property of the pasted text, not of your intent.
No, and trying to make it right first is the most common way to stall. Put the material in; the structure is proposed from what is actually there and changes as more arrives. A folder tree invented before the content exists describes work nobody has done yet.
vault_conventions.conventions.placement_classes
Say what is wrong and let the skill correct it. The files are readable on purpose and you may open any of them, but a correction made by hand is invisible to everything that maintains the file, and the next run can write over it. Stating the correction keeps one writer per file.
vault_conventions.per_folder
From ecosystem.yaml, contract 41 · core-skills 1.89.4 · read at build 2026-10-08
