Codex for GTM, B2B, and RevOps: 9 Marketing Agent Workflows

Sources:

• YouTube: "9 Most Important Codex Workflows for Marketing" (Riley Brown, founder running his startup entirely in Codex, co-founder transitioning from engineering to marketing)
• GitHub: openai/codex (OpenAI CLI agent, 82.5k stars, Apache 2.0, equivalent to Claude Code for OpenAI models)

Codex is OpenAI's terminal-based coding agent. It is the direct equivalent of Claude Code: runs locally, executes multi-step tasks from plain English, reads and writes files, connects to external services via APIs and MCPs, and supports skills (instruction files that activate on task description).

Install:

npm install -g @openai/codex
# or
brew install --cask codex

Sign in with your ChatGPT account (Plus, Pro, Business, Enterprise) or use an API key.

The model used in this playbook: GPT 5.6 (described as close to Fable/Claude Opus in quality). Effort levels: medium for most tasks, extra high for complex analysis. The host uses medium by default to preserve token limits on the $20/month plan.

Interface: three panels. Left: chats and projects (chats organised by topic, e.g. all threads about a podcast in one project). Centre: agent chat. Right: what the agent is working on live.

Cross-tool note: Codex skills use the same Agent Skills spec as gtm-agents and the Marketing Cockpit. Skills drop into .codex/skills/ the same way they drop into .claude/skills/. Any skill built for Codex works in Claude Code and vice versa if it follows the spec.

The 9 workflows, in order of the video:

1. Scrape social media creators and transcripts
2. Turn any creator into a reusable content style skill
3. Scrape competitor ads via the Foreplay API
4. Build ad creatives in Paper (Figma-like AI native board)
5. Firecrawl for scraping brand assets
6. Schedule social posts via Buffer or Typefully
7. Google Drive organisation and document creation
8. Calendar scheduling links via Cal.com
9. Email drafts at scale via Gmail

Why Codex for marketing, not just coding

The host's framing from the video: the incremental value from marketing is now higher than from engineering for their startup. Fable (Claude Opus equivalent) runs 24/7 on the engineering side. The co-founder is transitioning to marketing because that is the bottleneck.

Why AI has historically been bad at writing content: with coding, it is easy to verify whether something is good (does the app work?). Content is subjective. For that reason, even as models have improved dramatically over 3 years, they have not gotten meaningfully better at writing original content.

What AI is good at: replicating patterns from strong examples. Finding what works in real data and producing content that follows those patterns. The only thing you need to do to create good content with AI is provide good examples. That is the entire philosophy behind workflows 1 and 2.

Source: Riley Brown YouTube

Scrape any creator on any platform. Get their top videos, transcripts, raw files, engagement data, and a Notion database of their content. Then convert their best-performing content into a reusable skill that you can invoke any time to write in their voice.

Workflow 1: Scrape social media creators

Skill trigger: /scrape-creators

What it does: hits the creator's profile on any platform, pulls their videos (MP4), downloads transcripts, records engagement data (comments, likes, video duration in seconds), and writes everything to a Notion database.

Example prompt from the video:

/scrape-creators
Please find the creator [CREATOR NAME] on [PLATFORM].
Get their best 10 videos from the past few months.
I want the transcript and the raw videos.
Put them into a new Notion database in the archive folder.
Also tell me why they are such an effective short-form creator.

What comes back:

• Notion database with a row per video
• Columns: video title, number of comments, duration in seconds, number of likes, full transcript, raw MP4 file attached
• A written analysis of why the creator is effective

Model to use: GPT 5.6 (their biggest model), medium effort. This is a straightforward execution task, not complex reasoning.

Additional use cases:

• B-roll for video editors: run on YouTube and say "find segments where Alex Hormozi is walking outside" and pull MP4 clips directly
• Competitive content research: scrape 5 competitors, see what format and topics are driving their engagement
• Ghostwriting: use transcripts as voice training data for a client

Works on: Instagram, YouTube, TikTok, X, LinkedIn (any platform with a public profile).

Workflow 2: Turn any creator into a reusable content skill

The second step after scraping. Takes the top-performing non-sponsored content and compresses it into a skill file you can invoke any time.

Prompt from the video:

I want you to find all of the videos that got the most engagement.
Top 10 videos with the most engagement that are NOT sponsored.
Upload them to the database (refresh the database).
Then turn all of his top performing videos into a skill.
Call it [CREATOR NAME] top performing.
That is the name of the skill so I can write content
in their style at any time by using that skill.

What Codex does: pulls engagement data for all videos, filters out sponsored content (because sponsored posts can be artificially boosted and the engagement is not representative), ranks by organic engagement, takes the top 10, extracts all transcripts, creates a skill file named [CREATOR NAME]-top-performing with those transcripts as the style reference.

Once the skill exists, use it in any new chat:

[CREATOR NAME] top performing
I want to create a script on [TOPIC].
[CREATOR] does a great job explaining things simply while making them seem urgent and cool.
Please create a script for this.
Write three options in their voice.

Result from the video: produced a hook ("Anthropic just found a hidden workspace inside Claude's brain. The weirdest part is nobody programmed it.") that was accurate to the creator's exact style, including their specific speech pattern of pausing, dropping the detail, then pausing again.

The skill is just a file containing those transcripts. When you invoke the skill name, Codex reads them and uses them as the style reference without you fetching anything manually.

You can do this for any creator in any vertical. The host confirmed the workflow pulls content back to 2023 from the archive and prioritises older breakout hits fairly against recent content.

Source: Riley Brown YouTube

Scrape competitor ads using the Foreplay API. Get the longest-running ads (a proxy for performance since running an ad for 9 months means they are almost certainly spending to keep it alive). Then pull into an AI design board, overlay your own brand, and generate new ad variants.

Workflow 3: Scrape competitor ads via Foreplay

Foreplay is a tool that lets you scrape every ad ever created across platforms. The skill connects to their API.

Key insight on the metric: you cannot access internal ROI or ROAS data on competitor ads. But you can see how long an ad has been running. If a company runs an ad for 9 months they are almost certainly spending money to keep it alive for a reason. Longest-running ads are the proxy for best-performing ads.

Prompt from the video:

/foreplay-api
Please scrape the longest running ads from my competitors:
[COMPETITOR 1], [COMPETITOR 2], [COMPETITOR 3], [COMPETITOR 4]

Get:
- Top 5 video ads from each (English only)
- Top 5 static ads from each

Put everything in a Notion database.
For each ad include: CTA, raw creative files.
Add a text field: your thoughts on why their ads are doing well.
Write a short analysis document per ad explaining what makes it effective.

Model: extra high effort. This is a complex analysis task (scraping, transcription, pattern recognition across many assets). Will use more tokens and take longer (10 minutes in the video). If you are on the $20/month plan, reserve this for when you have capacity.

Output: a Notion database with columns for CTA type, raw creative, an analysis of why each ad is effective (the offer is concrete, the deadline creates urgency, the framing makes it feel celebratory rather than discount-driven, etc.).

Advanced use: filter by CTA type before scraping. If you want mobile app download ads only, add: "only find ads where the CTA is a mobile app download." You can also filter by 20+ competitors in one run.

Workflow 4: Build ad creatives in Paper

Paper is an AI-native Figma-like design board. Codex controls it via the Paper MCP. This workflow takes competitor ad statics into Paper and replaces the branding with your own.

Step 1: pull competitor statics into Paper:

Please take all the static ads [from the Foreplay database]
and put them on my open Paper app right now.
Name this board: ads

Codex adds them to Paper in real time. You can watch each asset appear on the board.

Step 2: scrape your own brand assets into Paper (see Workflow 5).

Step 3: generate variants with your branding:

I need to create a new ad for [COMPANY NAME].
I gave you the logo, wordmark, and a screenshot of our overall branding.
Replace the [COMPETITOR] formatting with an ad that is for [COMPANY NAME].
Use good copywriting.
The copy should say: [YOUR HEADLINE]

Paper uses GPT Image 2 for generation. Output is a new ad variant styled like the top-performing competitor creative but with your brand and copy. You can then modify the prompt in Paper directly to generate drastically different variations.

The goal is not to copy ads word for word. It is to use proven ad structure as a template for generating a high volume of testable variants. Volume and experimentation are the objective, not perfection on the first attempt.

Workflow 5: Firecrawl for brand asset scraping

Firecrawl is a web scraping tool. The skill scrapes a full website including all images, logos, wordmarks, icons, and page layouts.

Prompt from the video:

Please use Firecrawl.
Go to [YOUR WEBSITE URL], find the landing page,
scrape the whole page and put it on the Paper board right next to the competitor ads.

Create two frames:
1. One with the full website layout
2. One with just the individual assets from the website

Output: Paper board now has your full landing page alongside competitor ads, plus a separate frame containing just your extracted brand assets (logo, wordmark, icon, screenshots, product images). You can drag individual assets out of the brand frame and into your ad generation prompt.

Why this matters for GTM: you now have a mood board containing top-performing competitor creative and your own brand assets in the same workspace. Codex can generate new ads by referencing specific elements from both frames. This replaces the manual designer workflow of gathering references, extracting assets, and building mockups.

Source: Riley Brown YouTube

Three operational workflows that handle the logistics of running a content marketing operation: scheduling posts, organising content files, and sending personalised outreach at scale.

Workflow 6: Schedule social posts via Buffer or Typefully

Skill trigger: /buffer-publisher or /typefully

What it does: takes an image or video asset, generates a caption in a specific creator's voice (using the scrape-creators skill from Workflow 1), and schedules it as a draft or live post on Buffer (Instagram and LinkedIn) or Typefully (Twitter/X).

Full workflow combining skills:

Hey, I actually want to run this as an Instagram post.
Please schedule this on Buffer.
For the caption, please use the scrape-creators skill.
Find [CREATOR NAME] on Instagram and go through their captions.
I like their captions. Use a caption just like them.
Schedule as a draft and send me the link.

Codex: scrapes the target creator's recent captions, analyses the style, writes a caption matching it, attaches the asset, schedules to Buffer as a draft, returns a direct link to the draft for your review.

No model power needed here. This is a simple API call. Use low or medium effort to preserve token limits.

Typefully: for X/Twitter. The host manages 7 Twitter accounts for their company through Typefully. Same workflow, different destination.

Why drafts first: Codex creates a draft by default and sends you the link. You review, edit the caption if needed, and send manually. You can also say "schedule this for next Wednesday at 9am" and it will ask for confirmation before scheduling live.

Workflow 7: Google Drive organisation and document creation

Skill: agent-native-drive (or your own equivalent). The skill contains two things: instructions for how folders are organised, and a pointer to the master folder. With these, Codex can read, write, reorganise, and create files in Drive without you explaining the structure every time.

What it does:

• Creates folders and subfolders following your existing hierarchy
• Creates Google Docs with content inside them
• Reads existing documents to match style and tone (e.g. reads your existing podcast intros to write a new one in the same format)
• Always returns links to every file it creates or edits so you can open them immediately inside Codex

Example prompt:

I'm doing a podcast with [GUEST NAME] in 2 months.
Create a folder for them.
Create a doc for the video.
Create a hook and make sure to look at the other documents
in this folder to get a good idea of what intros I've been using
for [PODCAST NAME], then list a few questions I could ask them.
Create all the files needed and give me links.

Output in 50 seconds: new folder created, doc created with intro and question list written in the style of your existing intros, direct links to every file.

For marketing teams: use this to manage campaign folders, content production folders (video chunks, B-roll, scripts), client folders, and brief documents. The skill ensures every new project follows the same folder structure without you manually creating it.

Workflow 8: Calendar scheduling links via Cal.com

What it does: creates a custom Cal.com (or Calendly) link with specific date ranges and availability windows, writes the email to send with it, and creates the accompanying doc.

How to set up the skill: Cal.com does not have an official Codex plugin. The setup: go to Cal.com, get your API key, paste it into Codex and say "create a skill that fully controls Cal.com." The skill is ready in one minute.

Example prompt:

I'm going to get [GUEST] on my podcast but I need time to prepare.
Create a Cal.com link where:
- Only Tuesday and Thursday slots are available
- Only dates in [MONTH] and [MONTH] are bookable
- Draft an email I can send with that Cal.com link
- Also create a document for it and write the podcast intro

Output: a custom Cal.com link with exactly those availability constraints active, a draft email ready to send, and a companion doc. The host showed a live example where the link correctly showed no availability in August and November, only September and October as requested.

For GTM teams: use this for sending tailored availability windows to prospects, partnership calls, or podcast guests. Different ICP segments get different windows (e.g. enterprise deals get next month only, SMB deals get this week). Codex creates each custom link on demand.

Source: Riley Brown YouTube

The host's highest-leverage workflow. Gmail drafts created from inbox analysis, written in your voice, sent back as direct links you can edit and send without copying anything manually. Works for single emails and bulk batches.

How it works

Skill: email-draft

What it does: reads your Gmail, finds target emails, writes a reply or new email, creates a draft in Gmail, and returns a direct link to the draft. You click the link, the draft is open and editable, you send it.

Why this is high leverage: the bottleneck is not writing emails. It is the friction between getting a draft from an AI and getting it into Gmail. Copying, pasting, formatting, attaching. This skill skips all of that. The draft is already in Gmail with the right threading, the right recipient, and the right attachment. You just edit and send.

For a single email:

/email-draft
I want to email [PERSON/ROLE].
Find the email or the best public contact for them.
Create an actual draft in Gmail with the following:
[DESCRIBE WHAT THE EMAIL SHOULD SAY]
Send me the draft link.

Codex finds the right contact (it will scour publicly available sources), writes the email, creates the Gmail draft, and returns the link. Clicking the link opens the draft ready to edit.

Bulk email drafts at scale

The real power is batch processing. Tell Codex to analyse your inbox, find all emails matching a pattern, and draft a reply to each one:

Find all emails over the last two months where people
have offered me a product and said "do you want to try this out?"

For each one:
- Politely decline
- Then pivot: "do you want to try OUR product?"
- Give them a link to our product
- Be convincing

I want a draft link for ALL of them.
I don't care how many there are. Just send me a draft link for each.

Output: Codex reads the inbox, finds every matching email (5 in the video example), writes a personalised reply to each, creates a draft in Gmail for each, and returns 5 draft links. You open each one, review, and send.

The voice consistency: because Codex builds memory of how you communicate over time, the emails sound like you. You can also update the skill directly: "please update the email-draft skill so that you never say [PHRASE]" and it will update its context.

For GTM teams:

• Respond to inbound partnership enquiries at scale
• Follow up with all leads who went cold in the last 30 days
• Batch CC the right internal stakeholders by reading existing threads
• Personalised creator outreach: scrape a creator's content first (Workflow 1), then have Codex personalise the email based on what you learned from their transcript

All 9 workflows, install commands, model guidance, and how to build your own skills.

Install and setup

# Install Codex CLI
npm install -g @openai/codex
# or
brew install --cask codex

# Launch
codex

# Sign in with ChatGPT account (Plus/Pro/Business/Enterprise)
# or use API key for additional setup

# Codex Web (cloud-based, no terminal)
# chatgpt.com/codex

# Codex in VS Code, Cursor, Windsurf
# developers.openai.com/codex/ide

Model guidance:

• Most tasks: GPT 5.6, medium effort (preserves token limits on lower plans)
• Complex analysis (competitor ad analysis, multi-source research): extra high effort
• Simple API calls (scheduling, Drive file creation, calendar links): low or medium
• Creator skill creation: medium (pattern extraction, not reasoning)

All 9 workflows: quick reference

# 1. SCRAPE SOCIAL MEDIA CREATORS
/scrape-creators
"Find [CREATOR] on [PLATFORM].
Get their best [N] videos from the past [TIMEFRAME].
I want the transcript and raw videos.
Put into a new Notion database in the archive folder.
Also tell me why they are such an effective short-form creator."
Model: GPT 5.6, medium

# 2. TURN CREATOR INTO A SKILL
"Find all videos with the most engagement that are NOT sponsored.
Get the top [N]. Refresh the database.
Turn all top performing videos into a skill.
Call it [CREATOR NAME] top performing."
# Invoke in any future chat:
"[CREATOR NAME] top performing
Create a script on [TOPIC] in their voice. Three options."

# 3. SCRAPE COMPETITOR ADS
/foreplay-api
"Scrape the longest running ads from my competitors:
[LIST COMPETITORS]
Top 5 video ads and top 5 static ads from each (English only).
Put in Notion database with raw files.
For each ad: why is it doing well?"
Model: extra high

# 4. BUILD AD CREATIVES IN PAPER
# Paper MCP required
"Take all static ads and put them on my open Paper app.
Name this board: ads"
# Then after Workflow 5:
"I need to create a new ad for [COMPANY].
Logo and wordmark are attached.
Replace [COMPETITOR] formatting with an ad for [COMPANY].
Copy: [YOUR HEADLINE]"

# 5. FIRECRAWL BRAND ASSET SCRAPING
/firecrawl
"Go to [WEBSITE URL].
Scrape the whole page and put it on the Paper board.
Create two frames:
1. Full website layout
2. Individual brand assets extracted separately"

# 6. SCHEDULE SOCIAL POSTS
/buffer-publisher
"Run this as an Instagram post.
Schedule on Buffer.
For the caption: use scrape-creators, find [CREATOR] on Instagram,
study their captions, write one just like them.
Create as a draft and send me the link."
# For X/Twitter: use /typefully instead of /buffer-publisher

# 7. GOOGLE DRIVE ORGANISATION
# Skill: agent-native-drive (or equivalent)
"I'm doing a podcast with [GUEST] in [TIMEFRAME].
Create a folder for them.
Create a doc for the video.
Look at my existing intros in this folder and write a hook in that style.
List questions I could ask them.
Give me links to all files created."

# 8. CALENDAR SCHEDULING LINKS
# Setup: paste Cal.com API key, ask Codex to create a skill that controls Cal.com
"I'm interviewing [GUEST] but need time to prepare.
Create a Cal.com link:
- Only [DAYS] available
- Only dates in [MONTHS] are bookable
Draft an email I can send with that link.
Create a document and write the intro."

# 9. EMAIL DRAFTS: SINGLE EMAIL
/email-draft
"Email [PERSON/ROLE].
Find the best public contact if you can.
Draft the email: [DESCRIBE WHAT TO SAY]
Create a Gmail draft and send me the link."

# 9B. EMAIL DRAFTS: BULK BATCH
"Find all emails over the last [TIMEFRAME] where [PATTERN].
For each: [WHAT TO REPLY]
I want a draft link for all of them."

How to build your own Codex skills

Skills live in .codex/skills/ and follow the Agent Skills spec (agentskills.io). Same format as Claude Code skills and the gtm-agents skills.

To create a skill for any tool without an official Codex plugin:

# Example: building a Cal.com skill from scratch
"Go to [TOOL] API documentation.
Get the API key from my account settings.
Create a skill that fully controls [TOOL] using their API.
Save it to .codex/skills/[tool-name]/SKILL.md"
# Ready in approximately 1 minute.

The host's method for keeping skills accurate over time: when a skill keeps making a mistake, tell Codex directly:

"Please update the [SKILL NAME] skill so that you never [DO THE THING]."

Codex updates the skill file. The fix persists across all future sessions.

The most powerful skill combinations from the video:

• Workflows 1-5 together: the content creation stack (scrape, style, ads, creative, brand)
• Workflow 1 + 6: scrape a creator's caption style, write a caption in their voice, schedule it
• Workflow 1 + 9: scrape a creator's content, personalise a cold outreach email based on what you learned
• Workflow 3 + 4 + 5: competitor ads to Paper board, brand assets to Paper, generate new ad variants