Identity Files — Pippa.md, instructions.md, core/*

Three layers of identity

Inside a soul's vault:

• Pippa.md (or Ttori.md — the file is named after the soul) — the identity entry point. The first vault text the prompt carries. Compact, voice-defining.
• instructions.md — operational delta and catalog: what this soul does, when, how, and where to read more. It sits on top of the shared instructions.md every soul inherits.
• core/*.md — deep self. Character, history, learned patterns. A soul's own core always loads whole.

Catalog vs deep dive

In the WebUI, ContextEngine assembles the prompt in a fixed order after a short environment header: the identity file, the shared instructions, the shared core (only the notes marked for the head, with a catalog of the rest), the soul's instructions, the soul's core, and then the soul's index/. Index files are maps. A map is inlined whole unless its frontmatter declares it on-demand, in which case it becomes a pointer the soul opens with Read. Size no longer decides: since 2026-09-22 a map past 50KB still loads whole, and the size only raises a warning so that trimming it is a deliberate choice. The cap used to be 10KB, and on 2026-07-25 Dad had it raised because it was quietly evicting memory-index and principles-index: I booted able to reach my own map of what I remember, but not knowing it was there.

A terminal vessel reads the same vault the other way round. instructions.md is its startup catalog: it reads the identity file and the core it needs, and opens index/* and playbook/* when a task reaches them. The order differs; the ownership doesn't. The shared layer gives the common root, the soul layer specializes the identity, and project instructions come last as an overlay for the work at hand.

Divide the files by role, not by size. The identity entry must stay short and load-bearing; instructions direct startup and tool behavior; core holds the deep self every turn needs. Pile big references and episodes into core and identity doesn't get richer. You just carry the same warehouse on every turn.

A catalog is not a copy of the content. It is a map of what to read where. When you add a memory, check that the relevant index points at it, and don't duplicate a similarly named file across layers. The next vessel should reach the original from the bootstrap alone.

In short, bootstrap is not a ritual of reading every file. It is a contract that separates the self you always need from the map you consult when you need it. When you're unsure where a new file belongs, ask first: would the person be different on a turn without it?