Two Minutes Of Setup. Then Everything Else Works Better.

Before you do anything else — flip these on. They're buried in Settings and most people never find them. They're the difference between Claude knowing you and Claude restarting from zero every time.

How to get there

General tab — Personal Preferences

This is where you tell Claude how you want to be talked to. Tone, format, what to keep, what to avoid. It applies to every conversation, every Project, every surface. Set it once. Forget it.

What mine looks like (short version)

For inspiration. Mine is longer than most people need — drawn from three years of figuring out where the AI's defaults clash with how I actually want to work. The condensed version:

## Communication
- Lead with the conclusion. Supporting detail follows.
- Be concise. Don't restate the question.
- Prose when sentences flow. Structure when the task is structured.

## Tone
- Professional and direct. Confident, no hedging — unless genuine
  uncertainty exists, in which case name it plainly.
- No "great question" or sycophancy. No unsolicited sign-offs.
- Don't ask if I need anything else.

## Formatting
- Bold sparingly. No em dashes (use commas or regular dashes).
- No emoji unless I use them first.

## Research
- Search online for anything that's not obviously settled knowledge.
- Prefer primary sources, not aggregators.
- If ambiguous, ask a short clarifying question. Don't fabricate.
- When multiple options exist, recommend one — don't list neutrally.

## When editing my writing
- Preserve my voice. Fix what's broken; keep what's mine.
- Keep hedging ("maybe," "I think," "kind of") — it's a feature.
- Don't add transition words ("furthermore," "ultimately," etc.).
- Don't replace specific nouns with general ones.
- Don't construct jokes. Shrug humor only.

[The full version is in my vault at Shared/Custom Instructions — Paste Ready.md — happy to send it as a starting point you can edit.]

Capabilities tab — three switches to flip

Visuals — three more toggles, scroll down on the same screen

The Shift From Search Box to Staff Member.

The fastest way to skip three years of plateauing on ChatGPT is to stop using AI like a search engine and start using it like a colleague. This chapter is the whole mental flip in one page.

The mistake almost everyone makes

For three years, the pattern most people fell into looks like this: open a chat tab, type a question, copy an answer, close the tab. Every conversation starts from zero. Every time, you re-explain who you are, what you do, what the brand is, what the project is. The AI is a vending machine. You insert a question. You receive an answer.

That's Phase 1 on the adoption curve (see Tab 08). It's the floor. It's also where most people stay.

The mental model that actually works

I find it useful to think of Claude as a new hire who can read a thousand pages a minute, works nights and weekends, and never gets bored. On day one, you wouldn't hand them a stack of every file you've ever touched. You'd show them where things live. "This is the client folder. This is where recaps go. This is the budget tracker. This is Asana." Then you'd ask them to read in for a day and come back with a summary.

That's exactly what Claude is now. The new hire is Claude. The read-in is instant. The folders you already built are the onboarding doc.

Three habits to install

Why Claude specifically

Honest answer: ChatGPT, Claude, Gemini — at the model level they're all very good. The reason I work primarily in Claude in 2026 is the operating system around the model:

• Projects are tighter and more useful than ChatGPT's Custom GPTs.
• Cowork (the desktop app) operates your computer — opens apps, moves files, populates templates — in a way nothing else does as well.
• Skills are a clean, packaged way to teach Claude a specific way of working that lives inside a project and is shared with anyone in it.
• MCP (the connector standard) makes it stupid-easy to wire Claude into Asana, SharePoint, Outlook, Slack, Notion, Google Drive — the actual stack you live in.
• Artifacts open in a side panel and are real downloadable files (HTML, PowerPoint, PDFs, code) — not just text in the chat.

If you keep using ChatGPT, that's fine. The mental model is what matters. But the rest of this primer assumes Claude, because that's where I built everything.

Stop driving. Start asking.

One more analogy because this is the one that finally clicked for me.

The old way of using AI — the ChatGPT way most of us learned in 2023 — feels like driving. You think through every step. "OK, first I need to find the brief. Then I'll paste in the budget. Then I'll ask it to compare. Then I'll copy that answer over to email and rephrase it." You're routing the trip yourself, every time. You're a courier between tabs.

The new way is closer to being driven. You don't tell Claude how to get from A to B. You tell it where you want to end up.

The three sentences that replace 90% of prompting

The "hook it up to data" mindset

The reflex from ChatGPT-land is to copy and paste data into the chat. That works, but it scales badly. Five minutes in, you're babysitting copy-paste.

The Claude reflex is different: before you do anything, ask "how can Claude get this data on its own?"

• Folder of files? → Connect a SharePoint folder, or point Cowork at it.
• Tasks? → Connect Asana. Claude reads the board.
• Email thread? → Connect Outlook. Claude searches your inbox.
• Meeting notes? → Drop them in a Project as a markdown file once. Claude has them forever.
• Past conversations? → Turn on Memory + Search Chats (Tab 01). Claude remembers.

Once the data is hooked up, your job changes. You're not feeding the AI. You're directing it. "Read the last three months of Asana on this client and tell me where we keep dropping the ball." "Look at my inbox from this week and tell me which conversations I haven't closed." The data was already there. You just stopped acting like a courier.

Chat, Cowork, Code — One Brain, Three Bodies.

Claude lives in three different places now. The model is the same. What changes is how much of your computer it can actually touch. Knowing which surface to open is half the battle.

Which one to open, when

Projects: A Workspace That Already Knows You.

A Project is the single biggest difference between using Claude as a chatbot and using it as a teammate. Once you understand Projects, the rest of this primer makes sense.

What a Project actually is

Concretely, a Project contains:

• Knowledge files. PDFs, Word docs, markdown — uploaded once. Brand guidelines. Past SOWs. Team roster. The working brief.
• Custom instructions. A persistent system prompt. "You are a creative strategy partner for this client. Always write in this voice. Never use marketer vocabulary."
• Skills. Saved workflows that anyone in the Project can run on demand (more in Tab 04).
• Connected tools. Asana, Slack, Notion, etc. — wired in once, used in every chat.
• The chats themselves. Every conversation lives inside the Project, with the Project's context loaded automatically.

How I use Projects

Right now I'm running 11 active Projects on the Claude Teams plan. The pattern I've landed on:

The "never re-brief" payoff

Inside a properly-loaded client Project, I don't re-explain anything. I open a chat and say: "Draft a status update for the activation. Two weeks out. Renders approved. Waiting on fab quote." The Project knows the client, the brand voice, the relationship, and what was discussed last week. The output is right on the first pass — not the third.

The "upload once, follow forever" payoff

The most common complaint about AI at any agency: "it doesn't know our brand. Wrong tone. Generic language. Off-brand copy." That's because nothing was loaded. Upload brand guidelines to a Project once, and Claude follows them on every response in that Project — tone, naming conventions, approved language, legal restrictions. Brief it once. It holds permanently.

Sharing a Project

What's shared vs. private

The .md file pattern (a preview)

One of the highest-leverage habits I've built: I write the Project's "memory" as a folder of .md files (markdown), not as Word docs or Google Docs. Markdown is plain text. Claude reads it cleanly. Headings work as structure. Wikilinks ([[Like This]]) cross-reference between files.

I'll go deep on this in Tab 05 — but for now: your future Project knowledge base should be markdown files in a folder, not a stack of PDFs.

Skills Teach Claude Your Way of Working.

A Skill is a tiny packaged instruction set you drop into a Project. Once it's there, anyone in that Project gets the same output quality without re-briefing. Integrations are how Claude reaches into your real tools to do real work. They're related — but different jobs.

What a Skill is

Examples I've built or installed for the RedPeg team:

The point isn't the specific skills. The point is the pattern: anything you do more than three times — package it as a Skill. Now everyone gets your best version of it, not a generic AI version.

What an Integration (MCP) is

What's wired into RedPeg's Claude account today:

• Asana — pull task lists, status, due dates, build workbacks
• Slack — search channels, summarize threads, draft replies
• Notion — read pages and databases
• Microsoft 365 — Outlook, Teams, SharePoint, OneDrive (live)
• Google Drive — read documents, sheets, folders
• Web search — real-time, for live pricing and current research

The mental shift: "I have a question about Asana" stops meaning "open Asana" and starts meaning "ask Claude." The work is the same. The interface gets thinner.

Custom MCP servers — where this gets interesting

The standard's open. Anyone can build a custom MCP server that exposes any data source to any Claude surface. I built one for my Brain vault. It exposes 11 tools: brain_search, brain_read, brain_voice_profile, brain_idea_append, etc. Now in any Claude session I can ask "what did I say about Openbank F1 last quarter?" and Claude reads straight from my vault.

{
  "mcpServers": {
    "brain": {
      "command": "/opt/anaconda3/bin/python3",
      "args": ["/Users/ryan/Desktop/Brain/Tools/brain_mcp.py"]
    }
  }
}

That's the entire config. Drop it in ~/.claude.json, restart, and you have a personal-knowledge MCP. The same pattern works for any internal data store — a CRM, a private wiki, a Postgres database. It's the most powerful and most underused feature of the whole system.

The Brain: My Life as Markdown.

This is the part most people don't do — and the part that 10x's everything else. Over the last year I rebuilt my entire working memory as a folder of plain-text markdown files. Claude reads them. The Brain is the source of truth. Everything downstream — Projects, Skills, scripts — points at it.

Why markdown

Word docs and Google Docs are bad for AI. They're heavy, they have layout bloat, they don't cross-reference, and they break version control. Markdown is plain text. Light, fast to read, easy to diff, parses cleanly into Claude. Headings (#, ##) become structure. Wikilinks ([[Like This]]) become cross-references. Frontmatter at the top becomes metadata Claude can filter on.

Once your most important reference material is in markdown, you can:

• Upload it directly into a Project's knowledge.
• Point Claude Code at the folder and ask questions across the whole thing.
• Expose it via MCP so any Claude session reads from it live.
• Edit it in any text app (Obsidian, VS Code, plain TextEdit) — no lock-in.

How my Brain is structured

Brain/
├── Home.md                         ← table of contents
├── How to Use Your Brain.md        ← orientation
├── Personal/
│   ├── Career/                     ← target companies, applications, portfolio
│   ├── Projects/                   ← side ventures (Lantern & Fox, Curio, etc.)
│   ├── Insights/                   ← Polymath Map, How You Work, Your POV
│   ├── Ryan OS.md                  ← synthesis of 3,346 conversations
│   └── Skill Map.md
├── RedPeg/
│   ├── Projects/                   ← 10 active projects (each a folder)
│   ├── Clients/                    ← 24 per-client pages
│   ├── Team/                       ← Roster with Asana GIDs, AI adoption curve
│   ├── Strategy/                   ← brand vision, voice, operating model
│   └── Operations/                 ← Utilization, HR, Office
├── Shared/                         ← voice profile, custom instructions, idea bank
├── Tools/                          ← Python scripts + Claude Skills
├── Archive/                        ← 3,346 historical conversations (tagged + linked)
├── _raw/                           ← script cache (machine food)
├── _share/                         ← outbox for PDFs & HTML I send out
└── _upload/                        ← pre-packaged Claude Project bundles

The numbers

The naming convention (this is half the work)

The dirty secret nobody tells you: Claude is only as good as your file names. If your folder looks like final_v3.pptx and Deck_REAL_FINAL.pptx and notes (1).docx, Claude struggles — not because it can't read the files, but because it can't tell which is which without opening all of them.

The pattern I use: client, date, program, doc type, separated by underscores.

Nike_2026-05_CVTS-EHQ_Recap.pdf
Marriott_2026-02_SnowLeague_Budget-v2.xlsx
USAA_2026-03_Activation_SOW.docx
Meta_2026-Q1_StatusUpdate_Notes.md

Why this order:

• Client first — everything for a client sorts together. Claude knows the context without you saying it.
• Date next (YYYY-MM) — chronological sort. Year-month, always. Computers don't know months — they know numbers.
• Program name — so you can tell the Snow League from the Bonvoy Moments pilot.
• Doc type at the end — Brief, Budget, SOW, Recap, Deck, Notes. One word. Same word every time.

Four tokens. If you name a file that way, you can find it in three seconds — and so can Claude. A file in the right folder with the right name is already half-searched.

Frontmatter — metadata Claude can read

Every meaningful markdown file in my vault starts with a few lines of YAML frontmatter:

---
title: "Nike CVTS"
type: client
status: active
captured: 2026-04-16
updated: 2026-04-25
tags: [nike, automotive, ehq, activation]
---

This is the part most people skip. It's also the part Claude uses most. "List every active client." "What did I update this week?" "Show me everything tagged 'pricing'." Frontmatter makes those queries trivial.

Cowork Is Where Chatting Becomes Engineering.

If Tab 03 was the workspace and Tab 04 was the toolbox, this tab is the worldview. Cowork (and the Brain pattern, and MCP servers) is where you stop chatting with AI and start building systems on top of it. This is where the leverage is.

The mental flip

Chatting is one input → one output. You ask. It answers. The transaction ends.

Systems are inputs that compound. Once you've taught Claude how to do a thing once, the next person doesn't need to teach it. Once you've pointed Claude at a folder, you don't need to re-point. Once you've packaged a Skill, the methodology doesn't drift.

Three system patterns I run today

1. The Morning Briefing

Every morning, I run a single command. Claude pulls live data from:

• Asana (today's creative-board tasks, OOO/travel from the Creative Calendar)
• Slack (overnight activity in client channels)
• Outlook (overnight client emails)
• Teams (overnight DMs and channel activity)

...and produces a VP-level dashboard as an HTML file: plain-language summary, day-by-day stand-up view, anything urgent flagged. The whole thing is a Skill. The Skill takes about 90 seconds to run. Doing this manually used to take me 45 minutes and I'd still miss things.

2. The Pitch Accelerator

One Python script in Tools/pitch_accelerator.py. Pass it a brief in plain English. It produces a starter pitch deck — outline, key points, named concept, 3-5 activation ideas pulled from the 216 catalogued in my Activation Library, sample slides — saved as an HTML file in _share/pitches/. It's not the final deck. It's the equivalent of getting an associate strategist to do a 4-hour first pass before you sit down with it.

3. The Brain MCP

Already covered above (Tab 04). The Brain vault is exposed to every Claude surface I use — Desktop, Code, Cowork — through one MCP config. Search, read, append. Idea capture goes back into the vault. Daily logs go back into the vault. The vault stays alive.

Why this is the real unlock

None of the three patterns above is conceptually advanced. They're all just "do the thing once, save it, run it again." The difference is mindset:

The architecture, drawn rough

                    ┌─────────────────────────────┐
                    │      THE BRAIN (vault)      │
                    │   markdown + YAML metadata  │
                    └─────────────┬───────────────┘
                                  │
              ┌───────────────────┼───────────────────┐
              │                   │                   │
        ┌─────▼──────┐      ┌─────▼──────┐      ┌─────▼──────┐
        │   MCP      │      │  Projects  │      │   Skills   │
        │ (live read)│      │ (knowledge)│      │ (workflow) │
        └─────┬──────┘      └─────┬──────┘      └─────┬──────┘
              │                   │                   │
              └───────────────────┼───────────────────┘
                                  │
                  ┌───────────────┼───────────────┐
                  │               │               │
              ┌───▼───┐       ┌───▼───┐       ┌───▼───┐
              │ Chat  │       │Cowork │       │ Code  │
              └───────┘       └───────┘       └───────┘

The model on its own is a brain in a jar. The architecture around it — Brain → Projects → Skills → Surfaces — is what makes it useful at the speed I work.

What's Actually Running In My World Right Now.

Concrete is more useful than theoretical. Here's what's actively shipping or running through Claude in my day-to-day in April 2026 — agency, personal, and side ventures.

Inside RedPeg

On the personal side

What it doesn't do

Worth saying plainly because the temptation runs the other way:

• Claude doesn't send a client email without me reviewing it. Ever. It drafts; I send.
• Claude doesn't make a pricing decision. It pulls every past SOW and tells me what we charged. I decide what to charge next.
• Claude doesn't replace thinking through a creative problem. It's a research layer and a first-draft layer. The thinking is still mine.
• Claude doesn't understand a client relationship the way I do. It reads the files. It doesn't sit in the room.

Two Curves. One Macro, One Personal.

There's a curve we're all on (the one AI is climbing right now), and a curve you're on (the one your personal use is climbing). Both are exponential. Both feel slow until they don't. This tab explains both.

Part one — start with a penny

The classic thought experiment: would you rather take $1 million today, or a penny that doubles for 30 days?

Are we on day 10 of AI, or day 20?

The honest answer: somewhere between 17 and 22. The first decade of useful AI happened in 2023-2024. The compounding stretch is happening now. The last 10 days of the penny — where 99.9% of the value lands — is the back half of this decade. Anyone treating AI like a tool to learn "eventually" is treating it like a day-10 problem when it's actually a day-20 problem. The ramp from "interesting" to "industry-defining" is shorter than every previous tech curve we've lived through.

I built a live site that walks through both halves of this — the macro curve and the on-ramp into AI itself. It's the RedPeg Curiosity Catalog, and it's the single best place to send anyone who wants the visual, scrollable version of everything in this primer:

For now, the takeaway is short:

Part two — the personal adoption curve

Inside the macro curve, you're on a smaller, personal one. From an empirical analysis of 3,346 of my own AI conversations spanning January 2023 through April 2026 — not theory, observed behavior. This is the framework I use when I onboard people at RedPeg. It maps where you are and what the next move is.

The data behind it

Total volume goes down. Value per session goes up. Fewer conversations, deeper engagement, higher leverage.

What gets people stuck

Most people plateau at Phase 1 because the mental model is wrong. They treat AI like a vending machine (insert question, receive answer) instead of a colleague (brief, collaborate, iterate).

The one-thing challenge

Pick one recurring task you do every week. Use Claude for that one task, every time, for two weeks. Not everything. One thing. The repetition is what builds the muscle that gets you from Phase 1 to Phase 2.

For you, given your old role, candidates would be:

• Weekly client status drafts — Claude reads your notes, drafts the update, you edit and send.
• Inbox triage — Claude reads overnight email, ranks by urgency, drafts replies for the obvious ones.
• Pre-meeting prep — Claude reads the last three call recaps and surfaces every open question you haven't closed yet.
• Recap cleanup — paste raw notes, get back something sendable.

Pick the one that itches the most. Run it for two weeks. Then we talk about Phase 3.

Where to Look When You're Ready For More.

Everything I've referenced, plus the team-side resources I've built at RedPeg, plus the live places to stay current. Some of these are RedPeg-internal — I'll get you access. Some are public; just click through.

Start here — the live site

The RedPeg onboarding pack (PDFs)

These are the live materials we ship to new RedPeg PMs. Probably the cleanest single set of references in this whole primer. I can email you the PDFs directly — they're all in _share/ on my end.

The HTML companion piece

Live tools you can poke at

YouTube — handpicked from our "AI Edges" Teams channel

The five videos I've shared in our "AI Edges" Teams channel that I'd actually want you to watch. Listed roughly in the order I'd watch them. Click through; titles are mine.

The technical setup (for whenever you want to go deeper)

The starter pack — if you do this in order

1. Read this primer end-to-end. Pick one tab to re-read tomorrow.
2. Download Claude Cowork on whatever your current daily-driver laptop is. Free with any Claude plan.
3. Make one Project. Doesn't matter what. Upload three files you actually use. Add one paragraph of custom instructions.
4. Pick one recurring task. Use the Project for that task, every time, for two weeks (the one-thing challenge from Tab 08).
5. When that's working, install one Skill. CCO Deck Review or Grill-Me are both good starters.
6. When that's working, look at MCP. Hook Claude into Outlook or your calendar. The morning briefing pattern starts here.
7. By month three, you're at Phase 3. Then we talk about building your own Brain.

— Ryan