Most people use AI like a search engine. I use it like a team. This class shows you the inside of my MEGA BRAIN, the system that lets me run a performance marketing agency mostly on autopilot. You get the agent team, the knowledge base structure, the digestive system that turns courses into permanent knowledge, and the skills that do the heavy lifting. Every section comes with a downloadable starter pack you can install today.
🧠 1. The Big Idea
most people use AI like google. they ask one question, get one answer, then forget about it. that's a waste.
what i do is different. i built a system where my AI has memory, has a team, has a brain full of knowledge, and gets smarter every day i feed it. when i ask a question, it pulls from 14,000+ chunks of structured knowledge, routes the question to the right specialized agent, and gives me an answer that's better than what most consultants would give me.
that's not science fiction. that's a folder of markdown files and 6 agent definitions. you can build the same thing today.
the goal of this class isn't to make you a programmer. it's to show you the architecture so you understand what's possible. then i'm giving you a starter pack with all of it. you customize for your business and you have your own brain in 30 minutes.
🏛️ 2. The Architecture
before we go into each part, here's how the pieces fit together.
your-agency/
├── CLAUDE.md ← master config (the "brain bootloader")
├── .claude/
│ ├── agents/ ← your AI team
│ │ ├── ceo.md
│ │ ├── cmo.md
│ │ ├── coo.md
│ │ ├── cso.md
│ │ ├── cfo.md
│ │ └── jarvis.md ← the orchestrator
│ └── skills/ ← reusable workflows
│ ├── mega-digestive-lite/
│ ├── firecrawl-research-lite/
│ └── copywriter-lite/
└── .brain/ ← your knowledge base
├── _INDEX.md
├── domains/ ← chunks by category
│ ├── marketing/
│ ├── operations/
│ ├── sales/
│ └── finance/
└── dna/ ← expert personas (Halbert, Schwartz, etc.)
4 things working together. config tells AI who you are. agents are your team. skills are reusable workflows. brain is your knowledge.
you don't need to build all of this on day 1. you start with the CLAUDE.md, you add agents as you need them, you grow the brain over time. this class shows you what each piece does and gives you a starter version.
⚡ 3. CLAUDE.md — Your Brain's Operating System
this is the file your AI reads first, every single session. it tells it who you are, what your business is, what your voice sounds like, what rules to follow.
without this file, your AI is a generic assistant. with it, your AI is YOUR assistant. think of it as the difference between hiring a temp from kelly services and hiring someone who has worked for you for 6 months.
📄
CLAUDE.md Starter Template
A pre-structured CLAUDE.md you fill in with your agency's name, niches, voice, and goals. 15 minutes to customize, lasts forever.
~2 KB · markdown file
↓ Download
👥 4. The Agent Team — Your AI C-Suite
agents are AI specialists. each one has its own personality, its own focus, its own decision framework. when i ask my CEO agent about a strategic decision, it thinks like a CEO. when i ask my CMO about an ad campaign, it thinks like a CMO. they're not the same agent with different prompts. they're built differently.
here are the 6 i ship in the starter pack.
👔
CEO
Strategy, prioritization, GO/NO-GO calls
Your strategic thinking partner. Cuts what doesn't work, scales what does. Use when you're stuck on a big decision and need someone to think 90 days out instead of 90 minutes.
📈
CMO
Funnels, ads, creative, ROAS
The agent you'll use the most. Diagnoses campaigns, writes ad copy, picks angles, makes scale decisions. Built for the Andromeda era of Meta.
⚙️
COO
Operations, SOPs, delivery
Turns every recurring task into a process. Use it to write client onboarding flows, delivery checklists, and SOPs that make your agency run without you.
💼
CSO
Sales, scripts, objections, pipeline
Writes discovery scripts, handles objections, builds proposals, optimizes your pipeline. Use before every sales call.
💰
CFO
P&L, margins, unit economics
Sees the agency through numbers. Tells you which clients are profitable, which are bleeding, and what to charge for max margin.
🤖
JARVIS
The orchestrator
The agent that picks which other agent should handle a question. You ask Jarvis. Jarvis routes to CEO, CMO, COO, etc., synthesizes their answers, and returns one decision-ready response.
👥
Agent Team Starter Pack
All 6 agents (CEO, CMO, COO, CSO, CFO, JARVIS) plus the CLAUDE.md template. Sanitized templates you customize for your agency.
~10 KB · 7 markdown files
↓ Download
🌀 5. MEGA DIGESTIVE — Turn Courses Into Permanent Knowledge
this is the system that makes your brain compound. you bought a course. you watched a webinar. you read a book. most of that knowledge evaporates within a week. the digestive system turns it into permanent, searchable knowledge that your agents use forever.
the full version i use has 10 phases. way too complex for day 1. so i built you a 3-phase lite version that does 80% of the work.
PHASE 1
INGEST
drop raw content (transcripts, PDFs, books, transcribed videos) into a folder. give them a clean filename like 2026-05-08_haynes_challenge-funnel.md. that's it.
PHASE 2
CHUNK
tell your AI to break each file into 4KB knowledge chunks with structured frontmatter (id, title, tags, source, related). each chunk covers ONE complete idea. now your knowledge is searchable forever.
PHASE 3
SYNTHESIZE
turn chunks into actionable artifacts. you get topic dossiers (a 2-page master doc on a topic), expert DNA personas (how Halbert thinks, ready to use), or decision trees (if X then Y rules). your agents now know what these experts know.
🔥
This is the real flex. Most people consume content. You'll digest it. After a year of doing this, your AI brain will know more about your niche than any single human alive. That's not exaggeration. That's compounding knowledge.
🌀
MEGA DIGESTIVE Lite
The 3-phase pipeline as an installable skill. Drop content in, get structured chunks out. Includes domain codes, chunk format, and example outputs.
~3 KB · 1 skill file
↓ Download
📚 6. The Knowledge Base — Your .brain/
this is where everything lives. organized by domain. each piece of knowledge is a markdown file with structured frontmatter. each file links to related files. it's basically obsidian for your agency.
the starter pack ships with example chunks (an Andromeda creative volume rule, a 7-day onboarding SOP, a Halbert DNA persona) so you have something to copy from when you write your first chunks.
.brain/
├── _INDEX.md ← table of contents
├── domains/
│ ├── marketing/ ← MKT-ADS-001.md, MKT-CHF-001.md, etc.
│ ├── operations/ ← OPS-ONB-001.md, OPS-SOP-001.md
│ ├── sales/ ← SAL-DIS-001.md
│ └── finance/ ← FIN-PRI-001.md
└── dna/
├── halbert.md ← Gary Halbert persona
└── [your-experts].md ← add yours as you digest them
📚
Knowledge Base Starter
The .brain/ folder structure with index, example chunks (marketing + operations), and a Halbert DNA persona to copy from.
~5 KB · scaffolding + 3 example files
↓ Download
🛠️ 7. The Skills — Your Reusable Workflows
skills are reusable workflows. think of them as plays in a playbook. when you trigger one, your AI runs through a specific process every time, the same way.
i'm shipping 3 of the most useful ones for an agency.
🔍
Firecrawl Research Lite
Voice-of-customer mining
Pull real customer language from Reddit, Amazon reviews, YouTube comments, competitor sites. Outputs pain language, dream outcomes, objections, all in real customer words. The work that separates good copy from AI-flavored copy.
✍️
Copywriter Lite
4-master direct response system
A 4-step process combining Schwartz (awareness levels), Georgi (RMBC research), Halbert (hook craft), and Bencivenga (validation). Includes the master prompt to paste into Claude.
🌀
MEGA Digestive Lite
Knowledge digestion (covered above)
Already in section 5. Turns raw content into structured knowledge chunks.
🛠️
Skills Starter Pack (all 3)
Firecrawl Research, Copywriter, and MEGA Digestive Lite as installable skill files. Drop into .claude/skills/ and they're ready to use.
~8 KB · 3 skill folders
↓ Download
🧬 8. Memory System — Your AI Stops Forgetting
ok so by default, claude has goldfish memory. you open a new chat tomorrow, it forgets everything you said today. that's useless for an agency where you're juggling 5+ clients, dozens of decisions, and ongoing campaigns.
claude-mem fixes this. it's a memory system that uses a vector database (chroma) to save important things from your sessions. next time you open claude, it pulls the relevant memories automatically. you ask "what's the deal with the med spa client we onboarded in march" and it pulls the memory and answers like it remembers. because it does.
WHAT IT REMEMBERS
Four memory types, organized properly
User — who you are, your role, your preferences. Saved once, applies forever.
Feedback — corrections you gave claude. "Don't use em dashes." "Stop suggesting refunds." Saves you from re-correcting the same thing every session.
Project — current state of work. Which client launched when, which campaign is winning, what the next decision is.
Reference — pointers to where info lives in external systems. "Client billing is in stripe." "Performance reports are in google drive."
WHY AGENCIES NEED IT
Without memory vs. with memory
Without: you re-explain your business every session. Claude forgets what worked for client A. Every onboarding starts from zero. Knowledge stays in your head, not the system.
With: claude remembers every client. Decisions compound. New team members get up to speed faster because the memory teaches them. Your agency gets smarter every week, not just every quarter.
🔑
The unlock: Memory + Knowledge Base + Agents = your AI gets smarter every single day you use it. After 6 months it knows your business better than your last hire did. That's not exaggeration. That's compounding.
🧬
Memory System Setup (claude-mem)
Install guide, the global CLAUDE.md memory rules to drop into ~/.claude/CLAUDE.md, and the search patterns that actually work (vs the ones that error out).
~4 KB · 2 markdown files
↓ Download
🔌 9. MCP Stack — The Tools That Make Claude Actually Useful
an MCP is basically a USB port for your AI. every MCP plugs in a new capability. without them, claude can think but can't DO anything outside the chat. with them, claude can run ads, scrape websites, automate browsers, pull real data on demand.
there are 3 MCPs every agency needs. install these 3 and you'll do work in 10 minutes that used to take days.
📊
Meta Ads MCP (the official one)
Run, audit, optimize Meta campaigns from claude
What it does: claude can pull ad performance data, audit campaigns, suggest optimizations, even build new campaigns. All without you opening Ads Manager.
Why agencies need it: the difference between claude saying "check your CPA" and claude saying "your CPA on campaign X is $42, target is $30, here's what to kill and what to scale." Real data, real decisions.
The big news: Meta launched the official MCP on April 29, 2026. 29 tools, OAuth-based (no Dev App headaches), works in Claude Desktop and Claude Code.
After install you can ask: "audit my account for the last 30 days" / "which campaigns have CPA above target this week?" / "show me my top 5 ads by ROAS in the [client] account."
🔍
Firecrawl MCP
Web scraping at scale
What it does: claude scrapes any website. one URL, multiple URLs, entire sites. Returns clean markdown ready to feed into research, copy, or competitive analysis.
Why agencies need it: voice-of-customer research used to take days. With Firecrawl, you scrape Reddit, Amazon, and competitor sites in minutes. Your ad copy stops sounding AI because you're feeding it real customer language.
After install you can ask: "scrape r/[subreddit] top of all time and pull pain language" / "scrape these 5 competitor landing pages and tell me what hooks they use" / "scrape Amazon reviews on [product URL] and categorize objections."
🎭
Playwright MCP
Browser automation, real clicks
What it does: claude controls a real browser. Click buttons, fill forms, take screenshots, navigate sites that block scrapers.
Why agencies need it: some things you can't do with HTTP scraping. Logging into Meta Ads Library and exporting data. Auto-creating Framer pages from Figma designs. Taking screenshot proof for client reports. Playwright handles all of it.
After install you can ask: "open meta ad library, search [brand], take screenshots of top 10 active ads" / "go to my Framer site, log in, create a new page from this design" / "test the checkout flow on [client site] and screenshot any errors."
🔥
The combo move: Meta Ads + Firecrawl + Playwright running together = "find competitor brands ranking for [keyword] in Meta Ad Library, pull their top 5 ads, scrape their landing pages, screenshot their checkout flow, compare to our client's setup." That's a $5K research project. Done in 10 minutes. By claude. While you do something else.
🔌
MCP Stack Config
Install commands for all 3 MCPs, the settings.json template that wires them all up at once, and example prompts to test each one after install.
~4 KB · 1 install guide + 1 settings.json
↓ Download
🪝 10. Hooks — Auto-Magic In The Background
hooks are scripts that fire automatically when claude does certain things. think of them as your AI's reflexes. they run without you asking. they keep your brain organized, validated, and synced.
without hooks, you have to remember to link new chunks, validate frontmatter, save important learnings to memory, commit changes to git so nothing gets lost. you'll forget. every time.
with hooks, all of that happens automatically. the brain stays organized. knowledge compounds without your discipline.
HOOK 1
Auto-Link Chunks
Triggers when: you create a new file in .brain/domains/
What it does: scans the new chunk, finds related chunks already in the brain, auto-suggests them for the related: field.
Why it matters: every chunk needs at least 1 link. Without this hook, you'll create orphan chunks that the brain can't find later.
HOOK 2
Validate Chunk Frontmatter
Triggers when: any .md file in .brain/ is saved
What it does: checks the file has all required fields (id, title, tags, date, related). If anything's missing, blocks the save and tells you what to fix.
Why it matters: garbage in, garbage out. This hook keeps your knowledge base clean from day 1.
HOOK 3
Auto-Save Learnings
Triggers when: the session ends
What it does: scans recent messages for keywords like "lesson", "pattern", "key insight" and dumps them to your memory inbox for review.
Why it matters: claude often discovers patterns mid-session. Without this hook, those patterns die when the session ends.
HOOK 4
Auto-Commit
Triggers when: claude finishes editing files and the session ends
What it does: runs git add . && git commit -m "auto-sync" so your brain always has version history.
Why it matters: if you ever need to roll back, you can. If your laptop dies, you don't lose 6 months of brain work.
⚡
The shift: once these hooks are running, your brain maintains itself. You write knowledge. The system organizes it. You commit nothing manually. The system commits for you. You forget to link chunks. The system reminds you. That's the difference between a knowledge base and a self-improving brain.
🪝
Hooks Pack (4 hooks ready to install)
All 4 hook scripts (validate-chunk, auto-link, auto-commit, auto-save-learnings) plus the settings.json snippet to wire them up. chmod +x them, drop the config, restart claude. Done.
~5 KB · 4 shell scripts + install guide
↓ Download
🏗️ 11. Compound Engineering — The 4-Phase Framework
most agency owners skip planning, jump to execution, then pay for it on the back end. compound engineering is a workflow for any task bigger than a day that forces you to invest where the leverage actually is.
the rule is simple. for any non-trivial task, allocate your time like this:
PLAN ████████████████ 40%
WORK ████████ 20%
REVIEW ████████ 20%
DOCUMENT ████████ 20%
PHASE 1 — 40%
PLAN
most people give planning 5%. they get the mission wrong, the scope wrong, the order wrong. then they spend 3x longer fixing it during work. when you put 40% into planning, you understand the actual problem before lifting a finger. you catch dead ends in advance. you write the success criteria so you know when you're done.
PHASE 2 — 20%
WORK
once the plan is right, work is fast. you're executing a clear path. no decisions to make mid-flight. no second-guessing. 20% feels short. it is. that's the point. if work takes longer than 20%, your plan was incomplete.
PHASE 3 — 20%
REVIEW
you almost never get it right on the first pass. review is where you catch what work missed. read your output cold. pretend you didn't write it. find the holes. fix them. don't skip this. don't speed-run it.
PHASE 4 — 20%
DOCUMENT
this is the part everyone skips. big mistake. documenting is where the work compounds. you write down what you learned, what worked, what didn't. next time you face a similar task, you read your own notes and skip 70% of the planning phase. skip documenting and you'll redo this same work in 3 months.
🎯
When to use it: new client onboarding, launching a new ad funnel, building a new SOP, hiring decisions, pricing changes. Any project longer than 1 day. When NOT to use it: replying to an email, quick fixes, 30-min tasks. Match the framework to the size of the task.
🏗️
Compound Engineering Card
The 4-phase framework as a printable reference. Includes the exact prompts to give claude in each phase, anti-patterns to avoid, and a real onboarding example showing the framework in action.
~3 KB · 1 reference card
↓ Download
🎯 12. Ralph Wiggum — The 8-Pass Copy Loop
most AI copy is 6/10 quality. good enough to look like work, not good enough to convert. the ralph wiggum technique forces the same copy through 8 framework passes. after each pass, the copy gets stronger. by pass 8, it's 9.5+.
it's named after ralph wiggum from the simpsons, who repeats things until they stick. same energy.
you wrote a sales email. it looks fine. you ship it. it does 1.2% conversion. you should have run it through ralph. same email, after ralph: 3.8% conversion. that's the difference between "good enough" copy and copy that actually pays you.
PASS 1
Schwartz Awareness Check
Is this written for the right awareness level? Are we leading with the problem (unaware) or the offer (most-aware)? Mismatch kills conversion.
PASS 2
Hook Strength Check
Does the first line stop the scroll? Would you keep reading if you didn't have to?
PASS 3
Specificity Check
Replace every vague word. "Many" → "73%". "Soon" → "by Friday". "Growth" → "+$8K MRR".
PASS 4
Proof Check
Every claim needs a number, name, or example next to it. If you can't add proof, cut the claim.
PASS 5
Voice Check
Does this sound like a human or like AI? Em dashes. Triple structures. Corporate words like "leverage", "robust", "seamless". Kill all of it.
PASS 6
Pace Check
Mix sentence length. Use fragments. Use line breaks. White space is a weapon.
PASS 7
Cut Check
Cut 20% without losing meaning. Adjectives, qualifiers, throat-clearing intros, redundant transitions. All filler.
PASS 8
Cold Read Test
Read it fresh. Does it flow? Does it land? Does it close? If anything feels off, loop back to whichever pass missed.
⚡
The cost: 8-25 minutes per piece. Most people won't do this. They'll write copy in 5 minutes and ship it. That's why most copy is mediocre. You'll do it. Your copy will hit. Your campaigns will print.
🎯
Ralph Wiggum Lite Skill
The full 8-pass loop as an installable skill file. Includes the master prompt to paste into claude, the voice rules to enforce, and the quality bar to hit before shipping.
~3 KB · 1 skill file
↓ Download
📦 13. The Whole Thing — One Download
if you don't want to download each piece separately, just grab the full starter pack. everything you saw in this class is in there. CLAUDE.md, all 6 agents, all 3 skills, the .brain/ structure with example chunks, the install guide. one zip, one install, you're running.
🚀
MEGA BRAIN Starter — Ultimate Pack
The whole thing. CLAUDE.md + 6 agents + 4 skills (incl. Ralph Wiggum) + .brain/ scaffolding + Memory System setup + MCP Stack config + 4 Hooks + Compound Engineering card + INSTALL.md. Everything you saw in this class. One zip.
~45 KB · 25+ files
↓ Download All
⚙️ 14. Install in 30 Minutes
STEP 1
Unzip and rename
unzip mega-brain-starter.zip. rename the folder to your agency name.
STEP 2
Customize CLAUDE.md
open CLAUDE.md and replace every [BRACKET] with your real info. agency name, niches, voice, goals.
STEP 3
Customize the agents
open each agent file in .claude/agents/. replace [YOUR_AGENCY_NAME] with your real name. tweak missions to match your business.
STEP 4
Add your first chunk
read the example chunks in .brain/domains/. write ONE chunk of real knowledge from your business. that's your first compound.
STEP 5
Test it
open claude code in your project folder. ask: "read CLAUDE.md and tell me what you understand about my business." if it captures your business correctly, you're set.
🌱 15. The Compound
once this is running, the rule is simple. add one chunk per day to your brain.
in 30 days you have 30 pieces of permanent knowledge. in 90 days you have 90. in a year you have 365. each one makes every future answer better. each one makes your agents smarter. each one makes your business more defensible.
most agency owners don't have this. they have notion docs they never reread. they have webinars they watched once. their knowledge leaks every week.
you stop the leak. you compound.
— brez