ABC, always be cataloguing: folders, files, chat messages

If Protocol 03 says the agent is a folder, Protocol 08 says how you arrange those folders, and what you write into them, determines whether anything actually works.

This is the boring protocol. It is the protocol participants will try to skip. It is also the one that separates a setup that holds together for a year from a setup that turns into chaos after week three.

Folders: the canonical project layout

Every Infinite Leverage project has the same top-level shape. Same names. Same places. If you have seen one IL project, you can find your way around any IL project in 90 seconds.

• CLAUDE.md and AGENTS.md at the root
  Loaded by every agent on every run. CLAUDE.md is your project's brain stem; AGENTS.md is the cross-agent contract.
• Working Files/ for what you write with the PM agent
  product-plan.md, the nurture plan, the agent plan, the day-one reflection. The documents in your words, that the AI reads before it builds (Protocol 09).
• docs/ with subfolders
  docs/plans for the plans the AI builds from, docs/qa for every review, docs/handoff.md for the next person. Standardised so any agent knows where to look.
• brand/ for the design system
  Tokens, components, voice notes, examples. The folder every agent that writes UI reads first (Protocol 12).
• .claude/agents/ with one folder per agent
  Each agent gets its own folder with its role, skills, context, routines, guardrails, and Definition of Done. The folder is the agent (Protocol 03).
• src/ for code
  The actual application. App Router routes, components, libraries. Source code lives here and nowhere else.

Files: two rules that compound over time

These are the boring rules. They are also the ones that decide whether your project still makes sense to a fresh agent (or a fresh hire) six months from now.

• One name per concept
  Pick docs or documents. Pick product or pm. Pick epic or ticket. Then use that name everywhere, forever. Inconsistent naming costs every agent a search.
• Naming conventions live in CLAUDE.md
  Every agent loads CLAUDE.md on first read. Put your project's vocabulary there: file naming, branch naming, where things go. The conventions teach themselves.

Chat messages: what the conversation decided goes into a file

The most common way a project loses its mind: a decision made in a chat, three days ago, in a session that is now closed. The agent that made it has no memory of it. You half do. The next session rebuilds the thing you already rejected.

The fix is a habit, not a tool. Before you close a session, ask: what did we decide, what broke, what rule did we set? Write each one into the file where the next session will look. A rule goes into CLAUDE.md. A decision with rationale goes into docs/handoff.md. A thing that broke goes into the QA review. Then the chat can die.

• Rules go to CLAUDE.md
  "No em dashes." "Never push to main." "Status markers live in product-plan.md." Every agent loads it on every run.
• Decisions go to docs/handoff.md
  What you chose, why, and what you rejected. The layer most teams skip, and the one a fresh reader needs most (Protocol 18).
• Things that broke go to docs/qa
  The fix and the cause, next to the review that found it. The next build reads it before it starts.

Why this is the protocol people skip

Cataloguing is the boring protocol. People want to skip to building. The few who do not skip it are the ones whose projects are still legible after a year.

A line: a well-catalogued project ages like a library. A badly catalogued project ages like a junk drawer.

A fresh project with the IL shape, CLAUDE.md teaching every agent the conventions, the first agent ready to run, and one decision rescued from a chat into a file. Future-you and every future agent inherit all of it for free.