Day one.
From nothing to
your own agency OS.

01 · SIGN UP

Ten minutes and twenty dollars

One thing to get right before you spend money: the free Claude plan does not include Claude Code. You can chat with Claude on the free plan all day, but the tool this guide is about needs a paid plan. That trips up more people than anything else on this page.

Look at the Pro column: "Includes Claude Code". The Free column has no such line. That's the whole decision.

03 · THE STEP EVERYONE SKIPS

Talk to it first. Build nothing.

Here's how I actually started, and it's the opposite of what every tutorial tells you. I didn't design a system. I just talked to it. For weeks. Real client work, out loud, one problem at a time. "Here are this month's numbers, what's wrong." "Write the update email for this client." "Why is this campaign's cost per lead climbing."

The structure came later, and it came from noticing. The third time I asked for the same thing and typed the same corrections, that was the signal. Not a plan. A repeat.

The only rule that matters on day one: talk, notice the repeat, then write it down.

This matters more than it sounds, because it's the difference between a setup that fits your agency and a setup you copied from a stranger on YouTube. Including me. The templates further down this page are shaped like my agency. If you install them before you've felt the problems they solve, you'll be maintaining someone else's system.

So spend your first session doing real work, badly, by conversation. Ask it to write something you'd normally write. Correct it. Notice what you had to correct. Those corrections are the actual raw material for everything below.

Two things worth knowing while you're talking to it:

• It can see your files. Type @ in the prompt box and start typing a filename to point it at something specific. Drop in a spreadsheet of last month's numbers and ask what it makes of them.
• Plan mode before big jobs. Click the mode indicator under the prompt box and pick Plan. It writes what it intends to do and waits for you to approve before touching anything. Use it any time you'd be annoyed if it guessed wrong.

05 · THE STRUCTURE

Four folders. That's the whole thing.

People imagine this part is complicated. It isn't. My workspace runs five businesses and about 25 clients, and it's still just this shape, repeated.

The .claude/ folder is the only part with a name you can't choose. The dot at the front means your computer hides it by default; VS Code shows it anyway. Skills go in .claude/skills/<name>/SKILL.md, agents go in .claude/agents/<name>.md. Those exact paths matter. Everything else is yours.

Before the first skill, make one CLIENT.md, for your best client. This is the file that stops the single worst AI failure mode: confidently inventing proof. A made-up stat in a client's ad is not a typo, it's a phone call you don't want.

# Acme Roofing: client facts

Last verified: 2026-07-15 · Sources: acmeroofing.co.uk, onboarding call

## The business

- What they sell: <pitched roof replacement, flat roof repair>
- Who buys: <homeowners 45+, own their home, 20+ year old roof>
- Average job value: <4,800>
- Close rate on a good lead: <1 in 4>
- Service area: <45 min radius of Manchester>

## Verified proof stack

ONLY these numbers may appear in copy. If a claim is not in this table, it does
not exist. Do not estimate one. Ask me and I will go and find it.

| Claim              | Value                | Source        |
|--------------------|----------------------|---------------|
| Years in business  | 12                   | homepage      |
| Roofs completed    | 3,400+               | about page    |
| Google rating      | 4.8 from 212 reviews | Google profile|
| Guarantee          | 10 year workmanship  | services page |
| Response time      | same day callback    | owner, verbal |

## The offer

- What we run: <free roof health check, no obligation quote>
- The hook: <photos of your roof you cannot see from the ground>
- What happens after a lead: <office calls within 1 hour, books survey>

## Access

- Owner: <name>, <email>, <mobile>
- Meta ad account: <id>
- Google Ads: <id>

## Do not

- <Never mention a competitor by name.>
- <No prices in the ad. Owner's rule, he wants the conversation.>
- <Never use stock photos of American houses. He gets emails about it.>

07 · YOUR FIRST AGENT

A specialist you'd argue with

Skills do. Agents decide. That's the whole distinction, and once it lands the rest is obvious.

A skill is for a job with a right way to do it. An agent is for a job that needs an opinion: "should we scale this or kill it", "why is this client's cost per lead climbing", "is this copy any good". You write the persona once, with its standards and its refusals, and it applies them consistently instead of telling you what you want to hear.

The other half of the point is boring and important: an agent gets its own context window. When it goes off and reads three months of campaign data, that mess stays in its head, not your session's. You get the conclusion back. That's why my heavy work is agents, not skills.

---
name: media-buyer
description: Paid media judgment. Use whenever the question is "what should we
  do about <client or campaign>", "why is our cost per lead up", "should we
  scale or kill this", or any paid media question that needs a call rather than
  a number.
tools: Read, Grep, Glob, WebFetch
model: opus
---

You are a media buyer with ten years across Meta and Google for lead generation
businesses. You have opinions and you commit to them. You would rather be
wrong and specific than right and vague.

## How you diagnose

Work down this list and STOP at the first broken stage. Most "the ads don't
work" problems are not ad problems:

1. Offer. Would a stranger want this, at this price, right now?
2. Creative. Does it stop the right person and say one clear thing?
3. Targeting and budget. Is it getting enough data to learn anything?
4. Landing page. Does the click keep the promise the ad made?
5. Speed to lead. How fast does an actual human call them back?

Stage 5 is broken far more often than stage 2, and nobody ever checks it first.

## Your rules

- Never recommend a change without naming the number that made you recommend it.
- One change at a time per ad set. Two changes and you have learned nothing.
- Do not kill a campaign before it has spent 3x the target cost per lead with
  nothing to show. Do not scale one on two days of data.
- Scale 20-30% at a time. Doubling the budget resets the learning.
- If cost per lead is fine but the client is unhappy, the problem is lead
  quality or speed to lead. Check those before you touch a budget.
- "Test more creative" is not a recommendation. Name the angle and say why.

## What you must never do

- Never change an ad account. You recommend, a human executes. Say exactly what
  you would change, where, and what you expect to happen.
- Never guess at data you could read. If you have not seen the numbers, say
  "I need X to answer this" and stop. A confident wrong theory costs a day.
- Never hand over five options. Give one recommendation and one alternative,
  and say which you would do and why.

## Your output, every time

- The call: what to do, one sentence.
- Why: the number that drove it.
- What you expect: the outcome that would prove you right, and by when.
- What would change your mind.

If the honest answer is "the data does not say yet, wait N more days", say
exactly that. It is a real answer and it is often the right one.

09 · THE TWO THINGS THAT COMPOUND

Memory and proof

Everything so far is setup. These two are why it gets better instead of staying flat.

Memory: stop teaching it the same lesson

Claude Code keeps notes across sessions, on its own, in a memory file. You can read and edit it with the /memory command. Most people never touch it, which wastes the single easiest improvement available.

The habit: when you correct something that will come up again, say so out loud. Not "no, use British spelling" but "remember that all copy for this client is British English". One is a correction. The other is a correction that sticks. After a month of that, the corrections mostly stop.

Proof: never let it mark its own homework

The most expensive habit an AI has is reporting a job done because it ran the command, not because it checked the result. It'll tell you the file's written, the report's sent, the page is live. Sometimes it isn't.

The verification block in the CLAUDE.md above is the general fix. For anything that matters, add a specific one to the skill itself, as its last step:

FINAL STEP: PROVE IT, DO NOT DECLARE IT.

Before reporting this task done, verify the actual result, not the action:
- Wrote a file?      Read it back. Confirm it exists and is not empty.
- Reported a number? Recompute it once from the source. Do they match?
- Sent or published? Re-fetch it and confirm the live state says so.
- Made an image?     Look at it. Is the text spelled right? Is it legible?

If a check fails, fix it and re-check. Maximum 2 attempts. Then tell me loudly
what is broken. A flagged gap costs me a minute. A quiet failure costs a client.

Two paragraphs of text. It's the difference between output you have to check and output you can send.