Cody Schneider's marketing engineering crash course

1. 010:15

   The agent does the middle work

   Copy, ad images, uploads, keyword research, lead lists, blog posts, dashboards. The work between the idea and the result moves to a coding agent. You still pick the idea and judge the result.

   How
2. 022:59

   Data first, then tools

   Before the agent can market, it needs your numbers in one place: a pipeline that copies them in, and a warehouse that holds them. Everything else in the video sits on that.

   How
3. 035:30

   Read from one place, write to many

   The agent reads from the warehouse. It only calls Facebook or Google directly to change something: upload an ad, pause one, move one. Results flow back into the warehouse. That's the loop.

   How
4. 044:29

   One source can lie

   Ad platforms report their own wins. Follow the trail to signups and revenue in your own data before you trust a channel. On screen, his warehouse shows months of ad spend with zero signups.

   How
5. 057:35

   Eleven pieces of plumbing

   Pipeline, warehouse, a server that's always on, file storage, databases, timers, a place to host pages, logins, share links, git, and access to many APIs. You can build it from open parts or buy it bundled.

   How
6. 0614:53

   Keys live in a vault

   Give the agent a long-lasting ad token, but keep it in a secrets manager, not a file on your laptop. The agent uses it by name and never prints the value.

   How
7. 0718:35

   Several jobs at once

   He starts a job, and while it runs he starts the next one in another chat. Once a job works, he moves it to the cloud on a timer so it runs without him.

   How
8. 0816:15

   Read what comes back

   The agent swapped the image model, wrote ad copy he didn't ask for, flagged claims in its own ad scripts, and talked him out of 191 thin pages. All of it is in the replies.

   How

Marketing engineering

Marketing work done by telling a coding agent what you want, instead of clicking through each tool yourself.

GTM engineering

Go-to-market engineering. The same idea for sales: building lead lists, enriching them and sending outreach with code and agents.

Coding agent / harness

An AI that can run commands and edit files on a computer: Claude Code, Codex, Cursor, Hermes, a Grok bot. The harness is the app that runs it.

API

The door a service leaves open for programs. Facebook Ads, Google Ads and Search Console all have one.

Rate limit

The most calls an API will take in a given time. Go over it and you get blocked for a while, or worse.

Data pipeline

A tool that copies data from all your accounts (ads, analytics, CRM, payments) into one place, on a schedule.

Data warehouse

A database built to hold a lot of data and answer questions about it fast. The pipeline fills it; the agent asks it questions.

SQL

The language for asking a database questions. Coding agents write it well, so you don't have to.

Join

Lining up two tables by something they share, like a date or a campaign, so you can compare them.

Read vs write

Reading looks at data. Writing changes something: uploads an ad, pauses it, publishes a post.

Context window

How much text an AI can hold in mind at once. Millions of raw ad rows won't fit; a SQL answer will.

S3 bucket

Cheap online storage for files like images and videos. S3 is Amazon's; many others copy its format.

Postgres

A popular free database. Agents use small ones as notebooks: which ad used which prompt, which leads they found.

Cron job

A task that runs on a timer: every hour, every morning, every Monday.

Artifact

A page the agent builds for you, like a dashboard or a review board of ads, that lives at a link.

CLI

Command-line tool. A program you or your agent run by typing a command. Graphed's is installed with npm, the JavaScript package manager.

Secrets manager

A locked store for passwords and API keys. Programs fetch keys by name, so the key itself never sits in a file or a chat.

System user token

A Meta key that belongs to your business instead of a person. It can be set to never expire.

Ad set

A group of ads in Meta that share a budget, audience and schedule.

Pixel

A snippet on your site that tells Meta when someone who saw an ad signs up or buys.

GA4 / Search Console

Google Analytics 4 counts visits to your site. Search Console shows what people searched on Google before they found you.

Branded search

Someone searching your company's name. More of them usually means more people have heard of you.

Versus keywords

Searches like 'ClickHouse vs Postgres'. People typing these are choosing, which makes them good readers to reach.

Page-two keywords

Searches where you rank 11th to 20th on Google. Close to page one, so a little work can move them up.

Enrichment / waterfall

Enrichment fills in missing details like an email. A waterfall tries the cheapest provider first and only pays the next one for people still missing.

Email verifier

A service that checks an address will accept mail without sending one. Results: ok, catch-all (can't tell), or invalid.

UGC ad

User-generated-content style: a person talking to a phone camera about a product. Can be a real creator or an AI avatar.

Hook

The first line of an ad, meant to stop the scroll.

CMS

Content management system, where your blog posts live. His is Strapi.

He draws the first half of this on a whiteboard: sources into a warehouse, the warehouse into Claude Code, an arrow marked reading and one marked writing back to Facebook ads. The seven steps and their order are ours. 5:30

1. 01he says

   Make a developer app

   In Meta Business Settings, add an app. Name it anything.

2. 02he says

   Pick the ad use cases

   Create and manage ads with the Marketing API, and with Meta Ads Manager.

3. 03he says

   Connect your business

   Link the app to your business portfolio, confirm, create. Add a privacy policy link. About 10 minutes, he says.

4. 04he says

   Add a system user

   A user that belongs to the business, not to a person. This is what makes the token last.

5. 05he says

   Give it only what it needs

   Assign the ad accounts (and pages) it may touch. Nothing else.

6. 06his guide + Meta docs

   Generate the token

   His guide sets it to never expire, with every permission ticked. Meta allows that, but calls expiring tokens best practice. Tick only the ad permissions you need.

7. 07he says

   Put it in the vault

   Store it in a secrets manager and tell the agent its name. He pastes it into Claude Code and asks it to add it to his secrets.

His full walkthrough, with screenshots: Create a Facebook Ads developer token. Meta's own pages: system users and generating their tokens.

1. 116:15

   It swapped the image model, and said so

   Asked for GPT Image 2.5. The provider said "model not supported" to every name it tried, so it used GPT Image 2 and explained why. The narration still says 2.5.

2. 218:08

   It looked before it touched

   Before uploading, it said: "First I'll check the ad set without changing anything." It created the ad paused, and its script won't upload the same ad twice.

3. 322:10

   It found money with nothing to show

   April to August: no Facebook signups on almost $3k of spend. It's in the answer, and the video moves on without a word.

4. 426:30

   It flagged its own ad claims

   "The counts are estimates… Some hooks make claims about the product… Confirm the product really does each of these before the ads go live." It also fixed one: daily sync to hourly.

5. 528:00

   It talked him out of 191 pages

   He wanted "Google Ads vs Facebook Ads for [industry]" posts. It checked 573 versions across 191 industries; six had any searches. It suggested one page for the main query instead.

6. 626:50

   It said what it hadn't checked

   "I did not open the page in a browser, so I have not seen how it renders." Honest, and a reminder to look yourself.

## Marketing changes
- Before changing anything in an ad account, look first and tell me what's there.
- Create new ads and posts paused or as drafts. I turn them on.
- If you can't use the model or tool I named, stop and say what you'd use instead.
- List every claim an ad makes about the product, so I can check it.
- Before any paid run over [$5], tell me the cost and wait.
- Say what you didn't check (for example, "I didn't open the page").

## Marketing changes
- Before changing anything in an ad account, look first and tell me what's there.
- Create new ads and posts paused or as drafts. I turn them on.
- If you can't use the model or tool I named, stop and say what you'd use instead.
- List every claim an ad makes about the product, so I can check it.
- Before any paid run over [$5], tell me the cost and wait.
- Say what you didn't check (for example, "I didn't open the page").

Cheap per call, but an agent on a timer makes a lot of calls. The agent you use (Claude Code, Codex) is a separate bill. He says Graphed is free to start; Graphed's setup notes for agents say the CLI needs a paid plan. Check before you plan around it.

1. 01

   Pick one question you answer by hand

   The one you dread on Mondays. Did last week's ads make money? Which posts are close to page one? Write it down in one sentence.

2. 02

   Export the data yourself, once

   Download CSVs from Search Console (Performance → Export), GA4 and your ad account. Put them in one folder. No pipeline yet.

   Search Console export help
3. 03

   Open an agent in that folder and ask

   Claude Code, Codex, Cloudroom or bb. Paste the starter below. This is the read half of the loop, for free.

   bb
4. 04

   Make the folder a tiny warehouse

   Ask the agent to load the CSVs into DuckDB, a free database that's one file on your laptop, and answer with SQL from now on. Same idea as his warehouse, smaller. This step is our suggestion, not his.

   Load every CSV in this folder into a DuckDB file called marketing.duckdb, one table per file.
   From now on, answer my questions with SQL against it and show me the query.

   DuckDB
5. 05

   Add one write, and make it a pause

   Make a Meta system user token for one ad account, put it in a secrets manager, and let the agent pause ads by name. Pausing can be undone, which makes it the safe first write.

   The token, step by step
6. 06

   Put the copying and the question on a timer

   A pipeline (Airbyte is the open-source one he names) fills a warehouse every hour, and your question runs every morning on a schedule. Cloudroom and bb have scheduled automations; a cron job on a server works too.

   Airbyte
7. 07

   Or buy the whole stack

   His Graphed CLI bundles all eleven pieces. He gives Claude Code the npm page and says "install this". He calls it free to start; Graphed's own setup notes for agents say the CLI needs a paid plan, so check first. The SEO plugin is one more command.

   npm install -g @graphed-inc/cli
   graphed login
   graphed plugins add seo

   @graphed-inc/cli on npm

You're my marketing analyst. The CSV files in this folder are exports from:
- [Search Console: queries and pages]
- [GA4: sessions by source]
- [ad platform: spend and results by ad, by day]

Answer this: [your one question].

Rules:
- Read the files; don't guess. Say which file and columns each number comes from.
- Mark any period that's only partly in the data.
- If the sources disagree, show both numbers and say which you'd trust and why.
- Finish with the 3 changes you'd make, the one you'd do first, and what you'd need from me.
- Don't contact any platform or change anything. Read only.

You're my marketing analyst. The CSV files in this folder are exports from:
- [Search Console: queries and pages]
- [GA4: sessions by source]
- [ad platform: spend and results by ad, by day]

Answer this: [your one question].

Rules:
- Read the files; don't guess. Say which file and columns each number comes from.
- Mark any period that's only partly in the data.
- If the sources disagree, show both numbers and say which you'd trust and why.
- Finish with the 3 changes you'd make, the one you'd do first, and what you'd need from me.
- Don't contact any platform or change anything. Read only.

• Cody Schneider, “Marketing Engineering Crash Course with Claude Code” (posted Oct 14, 2026). The idea, the eleven-piece stack, the jobs and his prompts are his, retold in our words. Prompts marked as typed are read off the screen; the rest are from the captions. Numbers in “what came back” are read off the screen.
• Disclosure: the video is sponsored by Graphed, and Cody co-founded Graphed (per his LinkedIn). He calls it the sponsor on camera but doesn't say it's his. Every job in the demo runs through Graphed. Claims about Graphed's features, prices and results are his and Graphed's; nobody has checked them. See graphed.com/platform.
• Where screen and captions disagree, the screen won: the ad was made with GPT Image 2, not 2.5; Graphed's page says 750+ data sources, not 800+; its tools page says 232 tools, not about 250.
• His Facebook token guide is Graphed's, summarized. Meta's token rules are from Meta's developer docs.
• The seven-step loop, the nine-beat walkthrough, every fill-in prompt and starter, the job-market check, the dollar math, the setup order and the takes are FakeNerd's. Not sponsored by anyone on this page.