A fresh walkthrough just dropped, and it shows ChatGPT doing real work instead of just answering prompts. There are ten agents, one simple setup, and step 2 is the twist that makes it all work. The creator, marketing expert Grace Leung, shows 10 ChatGPT Work agents that take whole tasks off your plate, from research to proposals.
I’ve tried plenty of “AI agent” demos that fell apart after one slide. This one stuck with me because every use case ends with a finished deliverable, not a draft you still have to fix.
What’s new
ChatGPT Work is now an agent inside ChatGPT. The author explains the difference really clearly. When you prompt, you get an answer and the rest of the work is still yours. With Work, you give it a goal and some direction. It plans the steps, uses tools, and comes back with the finished thing.
It runs in two modes:
- Cloud (web or mobile): The agent keeps working even when your computer is off. The expert uses this for most tasks.
- Desktop app: She only switches to this when the agent needs files from a local folder.
Here are the 10 agents she shows:
- Storytelling agent: It turns reports and call notes into an interactive presentation on ChatGPT Sites. A custom “storyline first” skill makes every slide open with a key message instead of a dull title. You can hover over numbers, switch between leadership and team versions, and edit the copy right on the page.
- Web designer agent: It uses OpenAI’s free Product Design plugin plus Pinterest references to audit a dated dog grooming site for conversion problems. Then it offers three mockups and builds a live, working site with a booking form that really captures leads.
- Audience research agent: It runs sub-agents on Reddit, communities, and review sites at the same time. After about 35 minutes it delivered a 30+ page report built on real customer words, with clickable source threads and a new positioning idea.
- Data analyst agent: It looks into why demo requests dropped 30% across ads, traffic, and demo data. It found that a broken mobile conversion flow explained 95% of the loss and built a dashboard around that story.
- Lead gen agent: It goes back to lost leads, checks them with the LinkedIn plugin for buying signals, and scores them on ICP fit, pain, and timing. Then it drafts personalized emails in Gmail.
- Campaign planning agent: It connects to a Google Drive full of brand assets, pitches three campaign angles, and returns a brief plus six ad sizes in about 15 minutes.
- Competitive research agent: Three sub-agents click through competitor sites like real buyers. They produce a SWOT, a side-by-side tracker, and annotated funnel screenshots.
- Ghostwriter agent: It pulls 20 video transcripts with vidIQ and builds a voice profile and a reusable writing skill from them.
- Proposal agent: It turns a discovery call transcript into a branded proposal site where the prospect can pick a plan and accept on the spot.
- AI visibility agent: It reverse-engineers what ChatGPT and Google AI Mode say about one high-intent question. Then it writes a content brief with a fan-out map and section-by-section writing tasks.
The twist
Out of the box, the agent is like a smart new hire on day one. It’s capable, but you’d have to explain everything every single time.
The real leverage is what you set up around it. The original poster puts it in three layers:
- A project so it knows your brand.
- A skill so it knows how you do the work.
- A scheduled task so it shows up without being asked.
That’s what surprised me. The prompt barely matters. The setup does the heavy lifting.
Your mini-workflow
Here’s how to copy her approach for nearly any of these agents:
- 📁 Build a project. Load your brand facts, pricing, past work, and style into the project instructions. Connect your Google Drive if your assets live there.
- 🧠 Add a skill. Write down your process, standards, and favorite tools as a playbook. You can even ask the agent to save a successful run as a skill so it repeats it later.
- 🎯 Give a goal, not steps. Say what you want and point it at sources, then let it plan. For big research jobs, ask it to split the work across sub-agents.
- ⏰ Schedule it. The data analyst agent refreshes every Monday and sends an alert if demos fall more than 10%.
Pro tips from the video
- Clean your data first. If you’re mixing more than two data sources, the expert suggests a shared layer like Google Sheets that the agent can read.
- Build templates from your winners. Upload a campaign brief that performed well and turn it into a reusable template.
- One question at a time for AI search. She says going deep on one query beats spreading thin. Quality first.
- Expect blocks. Some sites may stop the cloud browser, so you might need to log in to get past them.
- Source material is everything. For the ghostwriter, the quality of your samples decides the quality of the voice. She also admits Claude still does voice work a little better.
- Verify, don’t blindly trust. The data agent gives you hypotheses, not final answers. Click into the source threads and check the numbers yourself.
The bigger idea I’m taking away is simple. Put whatever stays the same inside a project, attach only what changes, and ask for the finished asset. Freelancers can use that for statements of work, in-house marketers for budget requests, and consultants for audits.
Watch it in action
The live demos make this click much faster than any summary can, especially the website build and the self-serve proposal. 🚀 Check out the full video to see all 10 agents at work and grab the free agent skill pack she mentions.