Memory · 01 — The Codex

1. voice.md — how you sound

What it holds. The patterns, cadences, words, and bans that define your written voice. Not a brand-style guide (too abstract). A concrete description of how you write so Claude can produce text that lands as yours.

Template structure:

---
type: voice-profile
owner: [your name]
last-reviewed: [date]
---

# Voice Profile

## Sentence rhythm
- typical sentence length: [X words]
- short sentences for: [emphasis / one-liners / closing]
- long sentences for: [argument / context / qualification]

## Vocabulary preferences
- [3–7 words you use deliberately]

## Anti-patterns (never do this)
- [3–7 things you do NOT do — e.g. em dashes, "Excited to share", forced curiosity hooks]

## Verbatim examples
- [3–5 sentences pulled from real recent writing that exemplify the voice]

## Voice fingerprint (paste at top of any Claude prompt)
[3 lines that distill the above into a copy-pasteable header]

Prompts to extract it:

1. “Pull my last 10 LinkedIn posts (or recent emails / proposals / Slack messages). Don’t summarize them. List every distinct sentence pattern you see, every vocabulary preference, and every anti-pattern. Output as bullet lists I can edit.”
2. “Now write a 3-line ‘voice fingerprint’ I could paste at the top of any prompt to enforce this voice.”

Common mistake. People skip the verbatim examples section and write abstract descriptors (“direct, punchy, no fluff”). Claude can’t generate from abstractions. Three real sentences beat ten adjectives every time.

2. icp.md — who you serve

What it holds. A specific buyer profile. Not “small businesses.” A specific person at a specific company stage with a specific pain that you have specifically solved before.

Template structure:

---
type: icp-profile
owner: [your name]
last-reviewed: [date]
status: verified | hypothesis
---

# Ideal Customer Profile

## One-line definition
[Role + company stage + the specific pain you solve, in one sentence]

## Firmographics
- size: [X to Y people]
- stage: [revenue / EBITDA range]
- industry: [specific verticals or types of work]

## The buyer (not the user)
- title: [who signs the check]
- background: [what got them here]
- typical day: [what they spend time on]

## Verified emotional needs (from real prior clients)
| Need | Quote |
|---|---|
| Efficiency | "Help me save time and money through automations" |
| [...] | [verbatim from a real call] |

## What they tried that didn't work
- [3–5 alternatives they considered or used before you]

## Triggers (when they come to us)
- [3–5 specific moments that send them looking]

## Disqualifiers (when this is NOT the right ICP)
- [3–5 things that mean this is the wrong buyer]

Prompts to extract it:

1. “Read [paste the last 3 sales call transcripts or proposal docs]. For each, extract: who the buyer is, what triggered them, what they tried first, the verbatim emotional needs they expressed. Aggregate into a single ICP profile draft.”
2. “Compare this draft against [framework/framework.md ICP section]. Note where they agree and where they diverge. Flag divergences for me to resolve.”

Common mistake. Writing the ICP from who you wish bought from you instead of who actually has. The “verified emotional needs” section is the antidote — if you can’t quote a real prospect saying it, leave it out.

3. decisions.md — append-only decision log

What it holds. Every consequential business decision, dated, with rationale. New entries get appended; old ones never get edited. When a decision is superseded, the new entry references the old one.

Template structure:

---
type: decisions-log
owner: [your name]
append-only: true
---

# Decisions

## 2026-06-07 — Cortex is the brand name
**Context.** Old name "AI Operating System" was generic across consultancies. Bootcamp + Karpathy + Play Bigger research pointed at a single-word brand with category subtext.
**Decision.** Brand = Cortex. Category subtext stays "AI Operating System." Tagline ladder: *Your business, on autopilot. → Owner. Team. OS. → Memory. Senses. Habits. Learning.*
**Supersedes.** [2026-06-03 decision to keep "AI OS" as primary name]

## [next decision]
...

Prompt to backfill:

1. “List the last 30 days of consequential decisions I’ve made (look at git history, recent meeting notes in meetings/, and Slack saved items). For each, draft an entry in decisions.md format. I’ll edit before committing.”

Common mistake. People treat decisions.md like a to-do list. It is not. It is the record of choices already made, written so a future reader (you, or Claude in a session 6 months from now) can understand why the system is the way it is. If you find yourself writing future tense, you’re using it wrong.

4. sops/ — folder of repeatable processes

What it holds. One markdown file per repeatable process. Not every process — just the ones you do regularly enough that codifying saves time the third time you run them.

Template structure (per SOP):

---
type: sop
process: [name]
trigger: [what initiates this process]
owner: [who runs it]
last-run: [date]
---

# SOP: [Process Name]

## When to run this
[1–2 sentences describing the trigger]

## Inputs required
- [what must be on hand before starting]

## Steps
1. [step 1, concrete]
2. [step 2]
...

## Output
[what gets produced / where it lands]

## Pitfalls
- [3–5 things that go wrong and how to avoid them]

Start with these 3 SOPs first:

1. Onboarding a new client (your highest-stakes repeatable process)
2. Sending a weekly status update (your highest-frequency repeatable process)
3. Closing the books / running invoicing (your most error-prone repeatable process)

Common mistake. Writing SOPs in the abstract before you’ve run the process enough times to know what actually trips you up. If you’ve done it fewer than 3 times, you don’t have an SOP yet — you have a hypothesis.

5. meetings/ — meeting transcripts + classified notes

What it holds. One file per meeting. Granola (or Fathom, or Otter) generates the transcript. A skill like /meeting-cortex classifies and routes it. The Codex is where the classified note lands.

Naming convention:

meetings/
├── 2026-06-07_client-call_acme-corp.md
├── 2026-06-06_brainstorm_naming-the-cortex.md
└── 2026-06-05_customer-interview_jane-smith.md

Each file’s frontmatter:

---
type: meeting
date: 2026-06-07
classification: client-call | brainstorm | customer-interview | general
participants: [names]
related: [[icp.md]] [[decisions.md]] [[sops/onboarding.md]]
---

The auto-input pattern. Once /meeting-cortex is wired (Habits · 03), every Granola meeting becomes a file here within minutes. The Codex grows without you typing notes. This is the single highest-leverage habit in the whole playbook.

6. Wiring it together — wikilinks + index

What it holds. A top-level index.md (or use your existing README.md) that points to every file in the Codex. Each file uses [[wikilinks]] to reference related files.

Why wikilinks matter. When Claude reads voice.md and sees [[icp.md]], it knows to pull icp.md into the context too. The graph compounds: every file makes every other file more useful.

Template index.md:

# Codex Index

## Voice + identity
- [[voice.md]] — how I sound
- [[brand/]] — visual identity (if separate)

## Who we serve
- [[icp.md]] — ideal customer profile

## What we decide
- [[decisions.md]] — append-only decision log

## How we work
- [[sops/onboarding.md]]
- [[sops/weekly-status.md]]
- [[sops/invoicing.md]]

## What was said
- [[meetings/]] — meeting transcripts + classified notes

## How this connects to the rest of the OS
- [[../02-senses/]] — what touches the world
- [[../03-habits/]] — what runs on top of the Codex
- [[../04-learning/]] — how the Codex audits itself

Tooling: Obsidian renders [[wikilinks]] natively into a navigable graph. Plain markdown editors show them as text but Claude still uses them as routing hints. Either works.