⚡ Agency Automation

Automate Your Entire Agency with Claude Code

A vision-setting session covering what Claude Code actually is (a harness, not a wrapper), what it can do out of the box, real automations already running inside a live agency, and why the window to build this as a service is wide open right now.

📽 Workshop recording
~30 min read
Vision + Real examples
Informal live session
Section 01

The Moment We're In

This is the ChatGPT 2 moment. Not said lightly. The first ChatGPT moment was the mass public realization that AI could hold a conversation. This is the moment the same thing happens with AI that can take action — read your files, write your code, run your workflows, and operate your tools while you're not there.

🔥
The window is open now. Businesses don't have this yet. The consultants who specialize in this while it's still an emerging offer — who build the delivery playbooks, the reusable module libraries, the niche expertise — will have defensible positions when the platforms commoditize the simpler versions. The rush is on.

What's changing isn't just the technology. It's the business model math. When you can build faster, test faster, and deliver faster than a traditional team — performance-based pricing becomes viable, one developer can manage 5–10 client projects simultaneously, and an agency's cost structure looks fundamentally different.

🏗
Real number from the room: One agency's leader said he thought 50 clients would mean total burnout for his current team. With agentic systems running, he calculated he could reduce to 3 people and scale to 100 clients. Not a future projection — a current assessment based on systems already running.
Section 02

Wrapper vs. Harness — Why It Matters

The AI world went through a vocabulary evolution worth understanding, because it explains exactly what Claude Code is and why it's different from what came before.

The old world — Wrappers
A thin layer around a model
  • Took a base model and jerry-rigged functionality around its limitations — RAG systems, file search, custom prompting bolted on top.
  • When the model provider released native updates (Assistants API, knowledge bases, code interpreter), it obsoleted most wrappers overnight.
  • Thin. Fragile. Built to work around the model, not with it.
  • OpenClaude = ripped the Claude model out of a billion-dollar harness and put it in a crappy replacement. No web search out of the box. Dead end.
vs
The new world — Harnesses
A substantial environment around the model
  • Claude Code is trained via reinforcement learning to know how and when to use tools — not a feature bolted on, a core capability of the model itself.
  • Exists within a persistent environment (your workspace folder) with access to files, terminal, browser, and sub-agents out of the box.
  • Sturdy. Scalable. Gets more powerful as you add layers — context, data, commands, scripts, skills — without hitting a ceiling.
  • Progress curve: slower start, unlimited ceiling. The more you build, the more powerful it becomes.
📉
On OpenClaude specifically: The progress curve with wrapper-style tools hits a plateau or gets worse as complexity grows — more data, more edge cases, more failure modes. Claude Code's curve goes the other way. Each layer you add compounds. The gap widens over time. If you're on a wrapper, switch now.
Section 03

What Claude Code Can Do — Out of the Box

Before you add a single custom command or skill, Claude Code ships with a set of native capabilities that most people don't fully use. These are what make it a harness rather than a chatbot — the ability to take real action in your environment.

📂
File system access
Read, write, create, delete, and move files anywhere you grant permission — including outside the workspace folder.
🌐
Web search
Search the web natively, without any MCP or extra setup. Research, validate, and pull live information mid-task.
🤖
Sub-agents
Spin up background agents to work in parallel. One agent researches while another builds — without you managing either.
💻
Terminal execution
Run scripts, shell commands, and code. Anything you can do in a terminal, Claude Code can do — and script to run automatically.
🔌
MCP integrations
Connect to any external platform via MCP servers — CRM, calendar, email, databases, APIs. The integration layer is unlimited.
Scheduled execution
Wire any workflow to cron jobs or triggers. Work runs while you sleep — data collection, reports, audits, briefings.
💡
The mindset shift: Stop thinking about what AI can help you do manually. Start thinking about what AI can do while you're not there. If you can describe a task, you can automate it. Anything can be automated — if you can think it and describe it, Claude Code can build it and run it.
Section 04

The Context Pyramid — Layers That Compound

An AIOS isn't built in one go. It's assembled in layers, each one making the AI more capable and more useful. The metaphor is a pyramid — each layer needs the one below it to be stable before you build on top. Skip layers and the system becomes unstable at scale.

The reason for building in this order: each session, Claude Code reads these layers in sequence to orient itself. Without the bottom layers in place, the upper layers don't have a foundation to stand on. Context is everything.

Real-time Data
Live numbers — YouTube, Google Analytics, Stripe, Bitly, CRM. The chef's kiss on top. Starts manual, becomes automated over time.
current-data.md
Strategy
Current priorities, quarterly goals, what success looks like. Teaches Claude how to think like a strategist aligned with your direction.
strategy.md
Personal Role
Who you are, your relationship to the business, what you're responsible for. Claude knows how to help you specifically.
personal-info.md
Business Context
What this business is, who it serves, how it operates. Before Claude can help, it needs to understand what it's helping with.
business-info.md
Workspace Orientation
What is this workspace, where is everything, what tools are available. The foundation. Loaded every session automatically.
CLAUDE.md
🏗
Why this order matters: CLAUDE.md is always loaded first. It reads the workspace structure and knows where everything is. Then /prime reads the context files. Each layer adds one more answer: what is this thing → what is the business → who am I in it → what are we doing → how are we tracking. Once all five layers are painted, you have a system that can think alongside you.
Section 05

Real Systems Built — Not Hypotheticals

These are systems built and running inside a live business during the weeks leading up to this workshop. Not prototypes — production. They exist to show that anything you can describe can be built and automated using this approach.

📡
AI Model Tracker — Daily Intelligence Brief
Runs every morning before you wake up. Delivered via Telegram.
The problem: Claude Code's training data lags. When you ask it to build something, it recommends whatever model it was trained on — often inferior, outdated, wrong for the use case. You kept having to tell it to search the web for the latest Gemini model identifier.

The solution: A workspace that maintains a live reference library of all frontier models, automatically updated when anything changes — so Claude Code always knows the best tool for any job without searching.
  • 1Pulls from LM Arena leaderboard API every morning via scheduled script
  • 2Compares current top models against local database. Takes daily snapshot for trend tracking.
  • 3If a new model appears at the top: research it (web search + YouTube sentiment + community reactions)
  • 4Updates workspace reference docs with model documentation, API snippets, and use-case guidance
  • 5Generates content ideas based on the new model and delivers morning brief to Telegram

Result: Always up to date on the AI landscape. Content ideas generated overnight. Never manually researching models again. A task that wouldn't have had time for — now done daily.
✂️
Automated Video Editing Pipeline
Rough cut + B-roll suggestions. Editor becomes an operator, not a grunt.
The workflow: Drop a raw video file into a folder, run a script, get a draft cut out the other side — plus timestamped B-roll suggestions based on content context.
  • 1Transcription: OpenAI Whisper transcribes the audio with precise timestamps
  • 2Cut detection: AI analyzes transcript for natural cut points — removes filler, dead air, restarts
  • 3Draft cut: ffmpeg assembles the rough cut. Editor receives a working draft, not raw footage.
  • 4B-roll suggestions: Second pass analyzes transcript with brand context. Generates timestamped notes: "you mentioned Lovable here — scroll of the Lovable interface?"
  • 5Editor reviews: Thumbs up / thumbs down on suggestions. Editor becomes an optimizer of the pipeline, not a worker in it.

Result: Editor output increases. Burnout decreases. Channel can scale past the previous output ceiling. Content quality stays high because the creative decisions remain human.
📊
Business Mission Control Dashboard
All data from all businesses in one place. Month-over-month at a glance.
The problem: Data was scattered across spreadsheets, Google Analytics, YouTube Studio, Bitly, school platforms, and Stripe. Couldn't get a full picture of the business without spending hours compiling. Couldn't catch problems early.

The solution: A local database (running on your machine, not a cloud service) that pulls from all sources daily and surfaces everything through a dashboard — from YouTube views to bookings to revenue.
  • 1Data collection scripts pull from YouTube API, Google Analytics, Bitly, school platform, Stripe daily
  • 2Snapshot approach: APIs often return totals, not deltas. Take daily snapshots and calculate the difference yourself.
  • 3Daily data audit validates numbers — cross-checks that figures look correct before surfacing them
  • 4Dashboard: HTML view of all key metrics. Month-to-date vs last month. One screen, full picture.

On the data warm-up period: Takes about 7–8 days before deltas are meaningful. The goal is 30 days of history so you can compare month-to-date against last month. "Am I green for the month?" — that's the primary question the dashboard answers.
Section 06

/explore — From Idea to Built, via Voice

The /explore command is the bridge between "I want to be able to do X" and an actual working system. Combine it with voice input (Whisper Flow or any voice-to-text tool) and you can describe an idea out loud and walk through a full build process — no typing required.

It's particularly valuable for non-technical people because it handles the research and scoping phases that usually require knowing what's technically possible. You describe the outcome; it figures out how.

1
Describe the idea
Speak or type what you want to be able to do. Rough is fine. The command is designed to ask clarifying questions.
2
Scoping dialog
Claude enters plan mode and asks questions back and forth. Gets clear on the actual scope and functionality you want.
3
Research phase
Searches the web and your codebase. Finds the best tools and APIs. Figures out how to build it given what already exists.
4
Build
Executes the plan. Uses your existing workspace context, capabilities, and data. Anything you've built before gets reused.
🎙
Voice + /explore = the non-technical path in. You don't need to know how to build it. You need to be able to describe it. Run /explore, turn on your voice input, and talk through what you want. The research phase handles technical feasibility. The build phase handles execution. Your job is to describe the outcome and review the result.

Examples that sound impossible but aren't: AI that edits your videos, a system that monitors competitors daily, a tool that drafts client proposals from call notes, a dashboard that shows all your business data in one place. The /explore command will tell you what's feasible and how to build it — usually in under an hour.

Section 07

The Agency Layer Cake

An AIOS isn't set up all at once. It's a layered build — each layer runs on top of the previous one. The business model stays at the center; the AI layers wrap around it, each one making the business more capable, more autonomous, and more scalable.

Core
Your Business Model
Agency, consulting, ecomm, roofing — anything. This is what creates value.
Layer 1
Context OS
CLAUDE.md + context files. Claude knows who you are and what you're doing.
Layer 2
Data OS
Local database + scripts. Real-time data from all your platforms feeding the system.
Layer 3
Commands & Workflows
Slash commands for recurring tasks. Daily brief, data audit, competitor analysis, proposals.
Layer 4
Automation & Delegation
Cron jobs and triggers. Work runs without you. Tasks delegated to scheduled agents.
Layer 5
Growth Initiatives
New projects and capabilities built on the freed bandwidth. Ideas executed faster than ever.
📐
Start with Layer 1, always. The instinct is to jump to automation. Resist it. Context OS first. Data OS second. Each layer must be stable before you build on top. You can't delegate tasks the system doesn't understand — and you can't get good outputs without good inputs. Layers, not leaps.
Section 08

The Agency Opportunity — Right Now

The agentic engineering systems running inside this agency show what's already possible — and the surface area of what can be automated across any business is enormous. Here's what changes when you can build at this speed and cost.

  • Performance-based pricing becomes viable. When your build cost drops to near zero and your delivery speed multiplies, you can offer "we'll do the audit for free and take a percentage of upside" — because you can actually make money at it. A model that wasn't viable for traditional agencies is now on the table.
  • One developer, 5–10 projects. With agentic engineering pipelines: spec it → feed it to the pipeline → it builds, tests, reviews, ships → dev reviews output. All their job becomes is understanding requirements, preparing specs, and reviewing results. One person handling what used to take a team.
  • Start with internal apps. Don't start by pitching clients. Start by building internal tools for your own agency — onboarding systems, client dashboards, proposal engines, reporting tools. You learn the methodology, build your template library, and have proof of delivery before you go external.
  • The spec is everything. The limiting factor isn't the AI's ability to build — it's your ability to specify what you want. If you can properly spec something (especially coming off a client audit with all their context), you can feed it into an engineering pipeline and it will build. Get good at writing specs.
  • Employees become operators. The video editor doesn't manually cut footage — they review AI-suggested cuts and give thumbs up/thumbs down. They operate the system and optimize the pipeline. The same pattern applies across the agency: people move up the value chain, away from grunt work, toward judgment and direction.
Doing the thing, teaching the thing, selling the tools to do the thing. The most defensible position is to be someone who actually uses this at scale themselves — and can show a client a live, working example of what you're proposing to build for them. Your own AIOS is your best case study.
Key Takeaways

What to Remember

Takeaway 01
Claude Code is a harness, not a wrapper
It's a substantial, trained environment around the model — with native tools, persistent context, and a compound growth curve. Wrapper-style tools plateau. This one compounds.
Takeaway 02
Anything can be automated
If you can describe it, you can build it. The limiting factor is your ability to specify what you want — not the AI's ability to execute. Stop asking "is this possible?" Start asking "how do I describe this?"
Takeaway 03
Context is everything — build it first
The pyramid must be built bottom-up. CLAUDE.md → business → role → strategy → data. Without solid context layers, every other layer underperforms. This is not optional setup. It's the foundation.
Takeaway 04
Start with internal tools
Build for your own agency first. Learn the methodology, create your reusable templates, and have real proof of delivery before you go to clients. Your AIOS is your best case study.
Takeaway 05
Snapshot early, trend later
Many APIs only give totals. Take daily snapshots from day one so you can calculate deltas. It takes 7–30 days before you have meaningful trend data. Start the clock now.
Takeaway 06
The window is real — act now
Businesses don't have this yet. The practitioners building niche playbooks and delivery track records now will be the ones with defensible positions when this becomes mainstream. The window is open. Use it.
Act on This

What to Do Next

  • If you haven't done Setup & Context OS: Do that before anything else. Without the context pyramid built, every system you add on top will underperform.
  • Pick one internal task you do 3+ times per week. Don't start with the most ambitious automation. Start with the most repetitive thing you do. Use /explore to describe it and walk through the build. Get one win on the board.
  • Set up daily data collection on day one. Even if you don't have a dashboard yet, start snapshotting. Once you have 30 days of history, month-over-month comparisons become the most valuable data you have. Start the clock now.
  • Study the agentic engineering pattern. Spec → pipeline → plan → build → test → review → ship. This is the methodology that allows one developer to manage 5–10 client projects. GitHub Issues as the orchestration layer. Learn it.
  • Get off OpenClaude or any wrapper-style tool. The ceiling is real and it comes fast. The sooner you're building on Claude Code's native harness, the sooner you're building on a system with no ceiling.
Continue Learning

Related Modules