Guide: ChatGPT Work Starter Pack

ChatGPT Work: Putting an Agent to Work Across Your Marketing and Ecommerce Stack

We've spent a lot of these issues on Claude Cowork — how to hand a whole job to an agent and get a finished deliverable back instead of a chat reply. What we haven't covered yet is its closest counterpart on the other side of the fence: ChatGPT Work. If your team lives in the ChatGPT ecosystem, this is the equivalent, and it's worth building into your processes now. It draws on the same agentic engine capability that powers Codex, ChatGPT's developer tool, which is a big part of why it's so capable.

This is a quick run-through of how to get started, plus five uses I'd reach for first. One useful shortcut in the meantime: almost any Cowork use case we've already covered transfers straight across — the two platforms are similar enough that the same principles apply.

Understanding ChatGPT Work

Regular ChatGPT gives you an answer and you go do the work. Work does the work and hands you the finished thing — a document, spreadsheet, deck, report, or web app — for review. It's an agent that acts across your apps and files, stays with a project as long as it needs to, and turns a goal into completed output.

It runs on GPT-5.6, and you use it differently from a normal prompt. Instead of "write me three subject lines," you describe an outcome: "build a competitive analysis of these five brands," or "process this month's supplier invoices into a clean spreadsheet." Work then breaks the job into steps and works through them.

In the desktop app it sits alongside two siblings: Chat, for fast conversational help, and Codex, for software development. Work is the one you'll reach for to get real output done.

The principle to set from the start — the same discipline we preach with Cowork — is to treat Work's output as a strong draft to review, not finished truth. It asks for approval before sensitive actions and cites its sources, and you decide what it can access. Keep a human in the loop on anything that sends, publishes, spends, or deletes.

The building blocks worth knowing:

• Projects with source folders. Every task lives in a dedicated project. On desktop you attach local folders Work can read and edit; with multiple folders you set a primary one, which is where it looks for instructions and configuration.

• Local and cloud modes. Local mode stays on your machine and files. Cloud mode keeps running after you close the app and syncs across web, mobile, and desktop.

• AGENTS.md. A short markdown file in a project's primary folder that Work reads first, every session. Store your folder structure, rules, and standing context there so it stops rediscovering things and repeating work.

• Plugins = apps + skills. Apps connect a tool (Gmail, Slack, Drive, your CRM) and tell Work where to work. Skills teach it how to work.

• Sites, built-in browser, and scheduled tasks. Deploy dashboards and internal apps by URL, let Work drive a browser to operate tools with no connector, and run jobs once, on a schedule, or when something changes.

What You'll Need

• A paid plan. Work is on Plus, Pro, Business, Enterprise, and Edu, and is rolling out gradually — so access depends on your plan and workspace. If you don't see Work or the folder option yet, the rollout may simply not have reached you.

• The desktop app. It's the fullest surface — local file access and the built-in browser only exist there.

• Your plugins connected. Connect the apps your team uses (email, Drive or SharePoint, Slack or Teams, calendar, CRM, ESP, analytics) and add skills-only plugins like Spreadsheets and Presentations. On Business, Enterprise, and Edu plans, most plugins are off by default until an admin enables them.

• A model tier in mind. GPT-5.6 comes in three tiers: Sol (flagship), Terra (balanced default), and Luna (fastest and cheapest). Start with Terra and move up to Sol for complex or high-stakes jobs.

The Core Loop

1. Create a project and attach only the folder(s) the task actually needs.

2. Describe the outcome, not the steps — the deliverable, the format, your constraints, and how you'll judge whether it's good.

3. Review the plan before it runs, and steer or stop it if you don't agree.

4. Let it work, then review. Don't quietly fix its mistakes — tell it what was wrong and have it correct until it's right.

5. Save repeatable work as a skill. This is where process before prompts pays off: you can only turn a process into a reliable skill once you've run it cleanly yourself.

6. Schedule it if it recurs, and keep approval gates on anything that sends, publishes, spends, or deletes.

Five Uses to Reach for First

I've picked five that map cleanly onto how ecommerce and marketing teams already work — one from each part of the toolkit. Each notes the feature it relies on, what to feed it, and a goal prompt to adapt.

1. Repurpose one long asset into channel variants. Local files + skill. Local. Put the source asset — an article, webinar transcript, or long post — in a project folder.

From this article, create a LinkedIn post, a short-form video script, and a newsletter blurb — all in our voice, each self-contained.

Give it one or two examples of your voice per channel, specify length limits, and ask for the strongest hook variant for each.

2. Product-review mining. Data analysis + local files. Feed it a review export for one or more products.

Cluster these reviews into themes, objections, and use cases. For each, give the frequency and two or three real customer phrases.

Ask it to separate positive from negative drivers, and to output the customer language verbatim — that's what feeds your copy and PDP work.

3. Competitor price and promo monitoring on a schedule. Built-in browser + scheduled task + email app. Give it competitor product URLs and connect your email.

Check these competitor pages, capture price and any promo into the tracker CSV, and each morning compare against the last run — email me only when something changes.

Start manual to confirm it captures the right fields, then schedule it, and ask for a short change summary in the email rather than the whole file.

4. Draft review and inbound replies for approval. Inbox or ESP app + approval step. Feed it the reviews or inbound messages, plus a few example replies in your tone.

Draft replies to these in our voice — acknowledge the issue, stay on-brand — but do not send. Leave them for me to approve.

Say "draft only, never send" explicitly, and give it an escalation rule so it flags anything sensitive rather than answering it.

5. Standardise PDP copy from a spec sheet. Custom skill. Give it a raw spec or attribute sheet plus 3–5 example descriptions in your format.

Write product descriptions from this spec sheet following our format and length rules, one per product.

Save it as a skill once it's right, and specify word count, tone, and any compliance wording that must appear or be avoided.

That's five to get moving with, but there are 23 in total — spanning content and creative, data and analysis, automation and monitoring, and ecommerce ops and enablement. The full run-down, with the inputs and build tips for every one, is in the complete SOP: [the rest is a free gift from me :) 

The through-line is the same one that makes Cowork work: describe the outcome, keep a human on the approval gates, and save the workflows that prove themselves as skills. Get one process running cleanly, then let Work carry it — that's where the time comes back.

• OpenAI launches GPT-6 Astra, claiming state of the art on computer use, coding, and science, saturating FrontierMath Tier 4 at 98% and ARC-AGI-3 at 99.9%. Rolls out to ChatGPT tiers and API this week.

• Jensen Huang calls GPT-6 Astra proof of AGI, but markets barely react, with Nvidia down 2% while OpenAI-linked stocks like CoreWeave jump 15%. Economists say real interest rates, not stock prices, are the real AGI signal.

• IAB raises its 2026 US ad spend forecast to 12.3% growth on Olympics and World Cup demand, as AI formats like Performance Max and Advantage+ hit 12% of spend, projected to reach 27% by 2030.

• Google rolls out Meridian GeoX worldwide, an open source tool letting advertisers run geo experiments to measure incrementality and strip out double counted conversions from click based attribution.

• Shopping ad impressions fall even as click through rates climb, and one analyst suspects Google reserves AI Overviews for lower intent searches while keeping ads on high intent queries.

📊 PlusDocs - Create custom templates for Google Slides.

🌐 Browser Use - We enable AI to control your browser.

🤖 Doclime - Get answers from your documents

⚙️ CapGo AI - Make market research effortlessly fast. Seamlessly gather vast web information into spreadsheets in seconds.

😍It was great, all I needed

🙂Good, not great

🙄It sucked