Source: raw/ChatGPT_Work_Just_Changed_How_You_Do_Marketing.md (https://www.youtube.com/watch?v=aRnVwkY4BJo) and raw/ChatGPT_Work_+_Dot_-_Build_your_AI_Marketing_Team_Agents_+_Skills.md (https://www.youtube.com/watch?v=BzS93V2zFTg), YouTube auto-captions fetched 2026-10-30. Creator: Grace Leung (youtube.com/@graceleungyl) Platform: YouTube Duration: 25:48 (launch walkthrough) · 15:58 (agent team build) Published: 2026-07-09 (launch walkthrough) · 2026-10-03 (agent team build) Skill level: Beginner (no coding; desktop app)

Grace Leung, an AI-for-marketing educator, covers ChatGPT Work twice: an unscripted launch-day walkthrough of marketing use cases, and a post-DevDay build of a marketing agent team run through Dot, OpenAI’s always-on assistant. Together they show what a non-technical marketer can do in ChatGPT Work: skills, plugins, browser control, hosted Sites and scheduled tasks. They also show where she still prefers Claude. The October video is HubSpot-sponsored.

Key Takeaways

  • ChatGPT Work is a “dedicated work agent living inside ChatGPT.” In the desktop app a toggle switches between Work and Codex. Leung’s rule: use Work “when the deliverable is the business outcome” (research, reports, decks, spreadsheets, websites) and Codex “when the deliverables is a software or program.” She says both “use much of the same underlying agent technology.”
  • Use the desktop app and a project folder. The desktop app can read and write your local file system. Pick an existing folder or create a project, and every output is saved there.
  • Skills come in three levels. The skills panel lists skills at the user-account, project and system level, plus recommended ones. System skills come prebuilt (image generation, skill installer, skill creator). You can upload an existing skill, such as her brand-design skill, and ask the agent to install it at the account level.
  • Quick chat, then Attach to Task. Quick chat handles brainstorming. Attach to Task packages the conversation’s context into a Work task, where you can pick a folder and tag a skill. In her demo this turned a deep-research report into an on-brand strategy deck. The deck followed her brand-design skill but left out the logo the skill defined.
  • Browser control runs through Chrome. Install the Chrome plugin in Work plus the “Codex for Chrome” extension, and the agent can search and scan sites. Her demo ran eight Google Shopping queries for a sectional-sofa client and wrote a title-pattern playbook (configuration, materials, benefit clusters, pricing and promotion). She found it “really fast compared to Claude.”
  • Scheduled tasks turn one-off work into recurring jobs. She asked it to rerun the shopping scan “every Friday” and update the playbook. The schedule panel lists every task with its model, reasoning level and repeat setting. Her other suggestions: a daily brief, a weekly review, a follow-up monitor.
  • Plugins connect real systems. Google Drive (find, upload, create, move files), Gmail, Slack, Notion, Calendar, Asana, Canva and Figma are named. A shared Drive with campaign briefs and brand guidelines is the context layer for team work.
  • Sites deploys a website on OpenAI’s servers. From a Labor Day campaign brief in Drive, the agent generated on-brand images with the built-in image model and deployed a campaign site, shareable from the Sites list. Faults she saw: no logo and an alignment issue. In October she adds that Sites has “its own built-in database for the lead capturing,” so a landing page needs no external server.
  • Template creator turns documents into skills. It converts an existing document, such as a client-onboarding brief, into a reusable skill. Given a messy folder of materials for a new client, the agent called that skill and produced a filled-in brief.
  • End-to-end workflows chain plugins. In one demo the agent read three weeks of posts from a Notion content calendar, generated visuals, uploaded them to Drive and updated the Notion entries. It skipped two posts; her fix is to give an explicit date range. In another it found a GA4 scheduled-report email in Gmail, read the attached CSV, made a branded PDF performance brief and replied on the same thread.
  • Dot is a named assistant “designed to take an ongoing responsibility.” It “has its own computer,” can remotely control your computer’s files with permission, and can use Work, Codex and your connected tools. You can rename it (hers is “Emily”), add it to Slack, give it an email address, or call it by voice. She says Dot is “only available for the ChatGPT Pro plan.”
  • Skills vs agents. Use a skill “when a task is repeatable and the process is super clear.” Use an agent “when a task needs judgment”: it interprets a goal, picks skills and owns an area. She defines agents by persona, focus scope and role only, because “the biggest value of the agent is the judgment.”
  • A coordinator Dot runs the team. She messaged “Emily” in Slack to coordinate two weeks of launch social content. Dot routed the request into the brand project and called the growth-strategist and content-creator agents. About 10 minutes later it returned a review PDF of six post drafts plus a zip with one folder per post, then scheduled the posts in Notion using a content-scheduling skill.
  • Add routing guidance to agents.md. Once you have agents and skills, ask ChatGPT to add “brief routing guidance” to the project instructions so it knows when to delegate to a specialist agent and when to use a skill directly.
  • Multi-agent runs are visible. A lead-magnet task run with past campaign data showed a plan, the skills it would use and a subagents tab. After about 15 minutes it had produced a guidebook PDF, a landing page with an embedded lead form tested locally, and a deployment to ChatGPT Sites.

The October 2026 agent-team build

  • Project setup. Use a local folder (demo brand: Lexora, a language-learning platform for working professionals) with a brand context file (what the brand does, target customers, brand style), marketing output folders and an agents.md “onboarding handbook.” Add the folder to a new ChatGPT project as a source. Leung offers a free setup kit that scaffolds this structure.
  • Team shape. A marketing team lead as the main agent, with specialists under it, each owning one area. The transcript lists growth strategist, content creator, creative designer, website designer, then “growth strategist” again. The repeat is likely an auto-caption error, and a later demo mentions an “analyst” reviewing results. Dot sits outside the team as an executive assistant.
  • Skill 1, ad creative. It takes a campaign brief, generates Instagram ad concepts and images, then adds logo and copy. Output used the brand colours, CTA buttons and logo treatment, all saved locally. She stayed in Work to build it: “you don’t need to switch to Codex.”
  • Skill 2, website builder. It came with a design-reference source file listing well-designed sites. She asked Dot, not the project, to build a one-page free-trial offer site on ChatGPT Sites. Dot routed the task to a new Codex thread and iterated with her. She says it took “less than 30 minutes.”
  • Scale. The same method grows to “a total of 15 marketing skills” covering strategy, content creation and website building. She says you can start with one or two agents instead of the full hierarchy.
  • Model note. She found “the Astra model is really good at making these type of diagrams and flowcharts” in the growth strategist’s strategy deck.

Try It

  1. Install one existing skill at the account level (a brand-design or deck skill) and run one real deliverable through Quick chat → Attach to Task. Check whether it used every asset the skill defines, such as the logo.
  2. Schedule a weekly competitor scan. Give the agent a query list, have browser control scan the results, write a playbook, then set a Friday rerun that records what changed.
  3. Turn your best brief into a template skill with the template creator, then point the skill at a messy new-client folder.
  4. Chain three plugins in one prompt (Notion → image generation → Drive). Give an explicit date range so posts are not skipped.
  5. If you have Pro, add Dot to Slack and give it one coordinating job across an existing project, such as two weeks of launch content. Leave it to report back.
  6. After three or more skills and agents exist, ask for routing guidance in agents.md before giving the team a multi-agent task.

Implementation

Tool/Service: ChatGPT desktop app, Work mode (Codex mode sits behind the same toggle); Dot for always-on delegation. Setup: Download the desktop app → create a project or pick a local folder → install skills (skills panel, or ask the agent) → connect plugins (Google Drive, Gmail, Notion, Slack, Chrome plus the Codex for Chrome extension) → optionally name Dot and add it to Slack or email. Cost: No prices are given in these videos. Leung says Dot needs the Pro plan; for current plan prices and Dot availability see OpenAI DevDay 2026. Integration notes: Created skills land in a hidden folder in the project (captioned as “.agent”), and agent files in a folder captioned as “Doc Codex.” The exact folder names are unconfirmed. Agents are invoked by tagging them in the project, or through Dot from Slack, email or a phone call.

How this compares to Claude Cowork

Leung’s own comparison (July 2026, from the launch video):

  • Verdict: “just use both.” “For now I can see is not much difference.”
  • Stability: she reports repeated Claude outages “in client meetings, in workshop,” and found ChatGPT Work “more smooth from every use case that I have tried.” She also cites Claude’s “pretty strict usage limit.”
  • Interface: Claude Cowork’s “three tabs” confused her, and she prefers ChatGPT’s minimized quick chat and its plugin grouping.
  • Skill storage: ChatGPT shows account, project and system levels. In Claude “there’s so many ways the agent can store skill” that non-technical users cannot tell where a skill is installed.
  • Where Claude wins: design. Claude “is actually very strong in design or generating really polished good looking mockups, websites, decks.”
  • Browser control: she found Claude’s “slow and not quite accurate” compared with ChatGPT’s.

Wiki context:

Open Questions

  • Dot plan availability. Leung says Pro only. OpenAI’s DevDay recap lists Pro and Business Premium in eligible markets (OpenAI DevDay 2026). Her claim is narrower, not contradictory.
  • Folder names. The skill and agent folder names “.agent” and “Doc Codex” are auto-caption readings. They may be .agents and .codex; unconfirmed.
  • Team roster. “Growth strategist” is listed twice; the fifth specialist is probably an analyst.
  • Sites access. The July demo site asked her to log in again when opened. Default Sites visibility and public-sharing settings are not shown in these videos.
  • Usage cost of Dot-coordinated multi-agent runs (about 10 to 15 minutes each) is not discussed.