AI Brain Starter Pack

Files are memory.

Chats end. Files persist. A small set of routed files gives your agent stable context about who you are, what the company knows, what is happening now, and how repeated work should run.

The point is not to collect everything. The point is to give each kind of information one home, then keep those homes current through real work.

One brain, multiple agents

The runtime is rented. The brain is owned.

Claude and Codex enter through different instruction files, then work from the same context, memory, tasks, history, and skills inside one folder you control.

Runtime A

Claude

CLAUDE.md
Routes Claude into the owned brain.

Owned layer

Your AI Brain folder

standing and current context
visible memory and history
one list of open tasks
canonical shared skills

Runtime B

Codex

AGENTS.md
Routes Codex into the same owned brain.

Use one canonical instruction file with a relative symlink or generated mirror. Hidden tool memory can cache the brain, but it must not become the only copy. Prove the setup by making one approved update through Claude and reading it back through Codex.

Read the visual guide

The two-level ladder

Build the personal loop before the company system.

Each level is a standalone copy-out package. Start small, prove it on real work, then add the company layer.

Level 1, you

Your personal memory

Give the agent context about you and route live work into three homes.

about-me.md for standing context
one area file for current state
debrief-history.log for history
TASKS.md for open tasks
one real debrief

Follow Level 1

Level 2, company

Your company brain

Mine durable company knowledge from real documents and connect it to a weekly operating layer.

brain/Company.md for durable context
People files for important relationships
operating files for live work
five shared, dual-mode skills
a scheduled weekly rhythm

Follow Level 2

The four parts of a working brain

Capture. File. Maintain. Ask.

Every brain in this pack, personal or company, does four jobs. A brain that does all four keeps compounding. A brain missing any one of them turns back into the stale wiki it replaced.

Capture

How work becomes raw material: debriefs, meeting notes, dictation, and the documents you already trust. Capture immediately, file later.

File

Where knowledge lives: plain files, one home per kind of information. Current state is replaced in place, history is appended.

Maintain

How the brain stays true: the debrief and reflect rhythm, plus a person approving changes before the working brain moves. Brains rot without a rhythm.

Ask

How answers come out: the agent reads the smallest useful slice at the start of every session, and people ask the brain instead of interrupting a colleague.

How the brain operates

Turn scattered knowledge into trusted context.

The four steps below are the write path: how raw material moves from Capture into File, with Review as the Maintain gate.

Evidence

Keep the original files, meeting notes, reports, and system records untouched.

Refine

Extract the useful facts, decisions, procedures, and preferences. Cite the source and flag conflicts.

Review

A person approves, corrects, or rejects the proposed knowledge before the working brain changes.

Approved brain

File the checked knowledge where the agent can retrieve the smallest useful slice for the next job.

Work creates new evidence. Debriefs capture what changed, approved lessons return to the brain, and the next job starts from checked facts.

The deployed structure

The repository is the catalogue. Your folder is the system.

You copy the level you need into a separate root folder. Prompts and examples help with setup. Your deployed folder holds the real working files.

Level 1 output

your-folder/
  about-me.md
  TASKS.md
  debrief-history.log
  one-area/
    CLAUDE.md
  skills/
    debrief/
    reflect/
    morning-sweep/
    weekly-update/
    meeting-notes/

Level 2 output

company-folder/
  CLAUDE.md
  AGENTS.md
  brain/
    Company.md
    People/
    _source-docs/
  operating-system/
    current-context.md
    GLOBAL_TASKS.md
    DECISIONS.md
    DEBRIEF_LOG.md
    REFLECTION_LOG.md
    WAITING_FOR.md
    WEEKLY_UPDATE.md
  skills/

What lives where

The company layer separates durable knowledge from live operating state, so the agent knows which file is authoritative for each kind of information.

CLAUDE.md and AGENTS.md: Two entry doors into the same brain. Make CLAUDE.md canonical, then point AGENTS.md to it with a relative symlink where supported, or keep a generated identical mirror with a drift check.
brain/: Slow-changing company knowledge. Company.md holds durable facts about the business, while People/ holds context about important working relationships.
brain/_source-docs/: Read-only copies of the documents used to build the brain. They are evidence, not the refined context the agent reads on every run.
operating-system/: The fast-changing layer for current priorities, open work, decisions, debriefs, waiting items, reflections, and the weekly rhythm.
current-context.md: The live state of play: what matters now, what is in flight, and what recently changed. Update it whenever the picture moves.
GLOBAL_TASKS.md: One authoritative list of open actions across the company. It is the control plane for today, this week, soon, and parked work.
DECISIONS.md: The durable record of decisions and their reasoning, so settled questions do not need to be reopened from memory.
DEBRIEF_LOG.md and REFLECTION_LOG.md: Debriefs capture what happened and changed. Reflections capture reusable lessons about how the company and the agent should work.
WAITING_FOR.md: Work that is blocked on another person or event, with enough context to follow it up instead of letting it disappear.
WEEKLY_UPDATE.md: The latest approved summary of progress, priorities, decisions, risks, and next actions.
skills/: Repeatable workflows for debriefs, reflections, morning sweeps, weekly updates, and meeting notes. Each skill defines how finished work is routed back into the right files.

The hard rules

Keep the agent useful and under human control.

The AI drafts, you send.No autonomous sending of email, messages, or posts.
Manual first, automate later.Prove a workflow is useful twice before scheduling it.
Capture immediately, file later.Get the thought down, then route it into its proper home.
Never overwrite the brain.Add the operating layer around source material and preserve the originals.

When you are ready to go deep.

Different builders arriving at the same idea from different directions: files, context, and a rhythm that keeps them current. Start with our own long-form walkthrough, then work through the rest.

The AI Brain Layer, by Works: The method behind this pack in long form: what a brain is, why files beat platforms, and how the layers fit together.
GBrain, by Garry Tan: An open-source company brain from the President of Y Combinator, MIT licensed, with a companion eval repo. The closest open build of the same idea.
How We Built Our Knowledge Base, by Cerebras: An engineering team's version of the same problem. Extract from where information lives instead of forcing one platform, and scope context so answers stay high-signal.
Everything You Need to Build a Company Brain, by Eric Siu: Skills, evals, and loops, building on his earlier 5-layer company brain. The best single argument for giving every skill a definition of done.
Context is King, by Alex Lieberman: Why the operator who manages context best gets the most out of AI, with a framework that maps almost one-to-one onto the files in this pack.
AI Strategy Should Be a Skill Library, by Hiten Shah: The case for codified, repeatable workflows as the asset you actually own. The skills folder here is this argument in practice.
AI Second Brains Decay, by Cole Medin: State versus event, and why brains rot without an update rhythm. The exact problem the debrief and reflect loop exists to solve.
How to Build an AI-Native Services Company, by Y Combinator: Eleven minutes on where this way of working is heading commercially.
Jeff Dean at Princeton: The Chief Scientist of Google DeepMind on one human directing a hundred agents, and why context is the bottleneck.
How Teams Actually Scale With AI, Tiago Forte and Rachel Woods: The human side: how teams adopt this without it collapsing into one enthusiast's side project.
How to Get Your Company AI-Pilled, by Geoff Charles: Ramp's VP of Product on driving adoption inside a real company.
The Ontology of the Company Brain, by Slite: A 149-person survey of the category: what people expect a brain to be, why only 17% have one that works, and teardowns of ten real architectures, the pattern in this pack among them.
WikiSkill: The academic entry. Benchmarked evidence that keeping raw logs, a markdown knowledge layer, and executable skills as separate layers beats every alternative tested.

Stop starting from zero.