Source: raw/ChatGPT_Work_-_10_AI_Agents_to_Automate_Your_Work_Right_Now.md (https://www.youtube.com/watch?v=JWdvTUZoN0s), YouTube auto-captions fetched 2026-10-30.
Creator: Grace Leung (youtube.com/@graceleungyl)
Platform: YouTube
Duration: 23:31
Published: 2026-10-16
Skill level: Beginner to intermediate (no coding; uses plugins, skills, subagents and scheduled tasks)
Grace Leung demos ten marketing “agents” built in ChatGPT Work. Each one is a project for brand context, a skill or plugin for the method, and sometimes a scheduled task, given a goal rather than step-by-step instructions. Most demos run in the cloud on chatgpt.com, so they keep going with the computer off. The desktop app is used only when the agent needs local files. Brands in the demos are illustrative, and the lead data is fake. The video is HubSpot-sponsored.
Key Takeaways
- Agent vs prompt. “When you prompt, you get an answer and the work is still on you. With work, you give it a goal, directions, and it plans the step, use tools, and come back with the finished deliverables.” Out of the box it is “a smart new hire on day one.” What makes it useful is a project (brand context), a skill (your method) and a scheduled task (runs unprompted).
- Two run modes. Cloud (web or mobile) keeps working when your computer is off. Local (desktop) is for files in a folder on your machine.
- Parallel subagents for wide research. In the audience-research and competitor demos she tells the agent to run one subagent per source or per competitor, “so each one focus on one source all at the same time.” Her reason: it is faster and goes deeper “than one agent skimming everything.”
- Save each working run as a skill. Audience research, the proposal and the voice profile are each saved as reusable skills. Next time you supply only the new brand or transcript.
- The principle behind the proposal agent: “you put what stay the same inside a project and only attach what changes and ask the agent for the finished assets.”
- Check the agent’s work. On the data analyst she says “you’re not just trusting everything it says”. She calls the project plan from the competitor agent “a bit overkill,” and the campaign assets and website are first drafts, not launch-ready.
The ten agents
| # | Agent | Inputs and tools | Output in the demo | Run time stated |
|---|---|---|---|---|
| 1 | Report / presentation | Project with brand context; reports and call notes; ChatGPT Sites plugin; custom “storyline first” skill | Interactive presentation on ChatGPT Sites. Every slide leads with a storyline hook and key message; hover numbers for detail; one click switches leadership and team versions; editable and downloadable | — |
| 2 | Web designer | Existing client site (dog grooming), Pinterest design references, OpenAI’s free Product Design plugin | Conversion audit, three image mockups (editable with comments), then a live working site on ChatGPT with a booking form that captured a test lead. Copy kept, design changed | — |
| 3 | Audience research | Goal plus high-level source directions (Reddit, communities, review sites); cloud browser; three subagents | Audience-insight report of 30+ pages built on customers’ exact wording, with links to source threads, messaging directions and a proposed repositioning of the landing page | 35+ min |
| 4 | Data analyst | Google Sheet joining demo requests, ad spend and traffic; data plugin | Interactive dashboard that opens with the conclusion (mobile conversion breakage explained “95% of the lost demo”), device breakdowns and alternative hypotheses. Scheduled to refresh every Monday and alert on a drop of more than 10% | ~15 min |
| 5 | Lead generation | Lost-lead data; free LinkedIn plugin; presentation template; Gmail | Flags contacts who changed jobs (warm leads at new companies), scores each lead on ICP fit, pain and timing urgency, finds lookalike accounts, writes a two-week plan and drafts personalized emails | — |
| 6 | Campaign planning | Google Team Drive of brand assets connected through project instructions; template creator plugin to build a campaign-brief template | Three campaign angles. After she picked one: a full brief plus six assets (feed, square, story) using real product shots from Drive, saved back to Drive | ~15 min |
| 7 | Competitive research | Three competitor URLs plus own site; explore “like a buyer”; three subagents | Per-competitor Drive folders with annotated funnel-stage captures, a SWOT file, a competitor tracker (target buyer, positioning, offers, pricing, lead magnet, site flow), funnel breakdown, change log and project plan | 20+ min |
| 8 | Ghostwriter / voice | vidIQ to pull her last 20 long-form transcripts (or a folder of transcripts); subagents | Voice profile (identity, speaking patterns, story structure, examples, pre-final checks) saved as a skill, then used in ChatGPT documents to draft a hook and intro | — |
| 9 | Proposal | Project holding pricing, past proposals and brand kit; Sites plugin; discovery-call transcript | PDF plus live-site proposal with the prospect’s name and logo, a situation-first opening, the most relevant case study, three selectable pricing plans and an accept button | — |
| 10 | AI visibility | Brand facts in the project; reusable “AI search content brief” template; one high-intent question; browser checks of ChatGPT and Google AI Mode | Captured AI answers plus a content brief: answer snapshot, a fan-out map inferred from visible answers, a target brand description, headline and evidence structure, original-data suggestions, and a section-by-section writing appendix | ~15 min |
Agent-by-agent notes
- Report agent. A good presentation is “a story with one clear message and proof behind it.” The skill exists to stop slides becoming a pretty summary. Teammates can edit the Sites version directly, so “the team never needs to come back to AI to make the changes.”
- Web designer. She uses Pinterest for design references. Reviewing the image mockups before the build lets you comment on design before any code exists. Remaining faults: “less AI looking images” are wanted, and motion graphics are missing.
- Audience research works for “almost work on any niche.” Its value is that findings are “based on real comments, real requests, real complaints about the competitor products,” and each claim can be checked against its source thread.
- Data analyst. If the analysis uses more than two data sources, she recommends cleaning them first into Google Sheets or a BI tool “as a shared data analysis layer for LLM.” In her view the agent is most useful for building hypotheses.
- Lead generation starts from lost deals (“people who already said maybe”) rather than scraping new names. In real use she would split the work into two steps or add ABM enrichment tools.
- Campaign planning tip: build the brief template from a past campaign that performed well.
- Ghostwriter. Pasting samples and saying “write like this” “only copies the surface”. Analysing a body of work captures structure and word choice. She has used Claude for this and says “Claude still does a little bit better job here.” Source quality matters most: “spend some more time gathering good sources.”
- Proposal applies to other documents: a freelancer’s statement of work, an in-house budget request, a consultant’s audit report.
- AI visibility. “Tracking alone doesn’t change the answer.” Reverse-engineer what AI already says, find the lines it quotes, and become the obvious answer. Do “one question at a time” because one question gives deeper insight, and put quality before quantity. The agent may be blocked when checking AI answers, and logging in can get around it.
Try It
- Set up one project per brand with brand facts and instructions before running any agent. Every demo depends on it.
- Run the audience-research agent on your niche with one subagent per source type. Read three source threads yourself before using any finding.
- Schedule the data-analyst dashboard to refresh weekly, with a threshold alert on your main conversion metric.
- Start lead gen from closed-lost deals. Score them on ICP fit, pain and timing, and look for contacts who changed jobs.
- Build a proposal skill: keep pricing, case studies and brand kit in the project, and attach only the new discovery-call transcript each time.
- For AI visibility, pick one high-intent question, capture what ChatGPT and Google AI Mode answer today, and turn the gap into a content brief. See AI SEO for the wiki’s citation research.
Implementation
Tool/Service: ChatGPT Work, cloud by default (chatgpt.com, web or mobile) and the desktop app for local files. Setup: Projects with instructions; plugins named in the video: ChatGPT Sites, Product Design (OpenAI, free), data, LinkedIn (free), presentation, Gmail, Google Drive, template creator. External: Pinterest, vidIQ. Cost: Not stated. The HubSpot-sponsored “marketer’s agent skill pack” (content research, marketing copywriting, marketing strategy, data visualization, landing-page audit) is offered free. Integration notes: The demos run from about 15 minutes to more than 35 minutes per agent. Subagents are requested in the prompt, not configured separately.
How this compares to Claude
- Source-stated: for voice-matched script drafting, Leung has used Claude and says Claude “still does a little bit better job here.”
- Wiki context: the project + skill + scheduled task combination matches the Claude Cowork setup of projects, skills and scheduled tasks covered in Claude Cowork for Marketing and Scheduled Tasks. Equivalent Claude skill packs for marketing and lead gen: Ben AI’s marketing skills and lead generation and outreach skills.
Open Questions
- Plan and usage cost of 15 to 35 minute cloud runs with three subagents is not stated.
- Caption uncertainties. “SWAT file” and “swap file” read as SWOT (high confidence). “Potential final queries” probably means fan-out queries (unconfirmed). “Session by session writing assignment” probably means section by section.
- Data plugin identity. The video says “the data plugin” without naming it.
- Lead-data privacy. She says the LinkedIn plugin uses “public info, so nothing is sensitive.” Combining it with CRM lost-lead data in a vendor cloud is not discussed.
Related
- ChatGPT Work for Marketing (Grace Leung) — her launch walkthrough and the Dot agent-team build.
- OpenAI DevDay 2026 — the platform changes behind these features.
- Codex Sites — the hosting product behind ChatGPT Sites.
- Claude Cowork for Marketing — the Claude-side equivalent.
- Marketing Agents That Run Facebook Ads (Cody Schneider) — another marketing-agent build.
- AI SEO — the citation research behind the AI-visibility agent.