Today's assistant features for marketers: projects, memory, deep research, files, voice and agents

0:00 Assistant features for marketers

If you last looked at AI assistants a year or two ago, you might still think of them as a chat box where you type a prompt and get a caption. That picture is badly out of date. Today's assistants have projects that hold a whole brand kit, memory that learns your preferences, deep research that writes cited competitor reports, file analysis that reads hundreds of reviews, voice for brainstorming, and connectors and agents that pull your files together. In this lecture you'll see how each feature maps to real marketing work, follow a marketer through a whole launch week, and watch me set up a brand project from scratch.

0:48 Why features beat prompts

Why do these features matter more than clever prompts? Because most bad AI marketing output isn't caused by a weak prompt. It's caused by missing context: the brand voice, the approved facts, the real customer language, the current competitor landscape. Features solve that systematically. A project holds your brand kit once and applies it to every chat. Deep research brings in current, cited information. File analysis brings in your actual data. Connectors bring in your team's documents. Once the context is in place, even simple prompts produce strong work.

1:27 The marketing studio analogy

Here's a way to picture it. Think of your assistant as a marketing studio. The project is the studio itself, one per brand or client. On the wall is the brand kit: voice guide, best past work, approved facts. There's a research desk, which is deep research, where someone investigates competitors and audiences and brings back a cited report. There's a data desk, which is file analysis, where someone reads your reviews and campaign exports. There's a rehearsal room, which is voice, where you practice pitches. And there's an assembly bench, which is connectors and agents, where someone gathers everything from your drives into a finished document. Same assistant. Different rooms for different jobs.

2:17 Projects and memory

Let's go room by room. The project. A strong marketing project has four parts. Instructions: audience, positioning, tone, words to use and avoid, approved and banned claims, disclosure rules and default formats. A brand kit: a one page voice guide, three to five best performing posts or emails, and a product sheet with approved facts and prices. Market context: personas and real customer questions and objections, anonymized. And a do not list: claims that need legal sign off, sensitive topics, competitors you must not mention. Then memory and custom instructions describe you, your role, spelling and format preferences. Keep client specifics in the project, not in global memory, so clients never mix.

3:06 Research, files and voice

Next, the research desk. Deep research runs many searches and writes a cited report in a few minutes: ideal for competitor positioning, audience questions and market entry briefs. Brief it with the decision it informs, the markets, the time frame and the sources you trust, and spot check five links. Then the data desk. Upload anonymized reviews, survey exports or campaign CSVs and ask for themes, example quotes and patterns. Three hundred reviews become five clear themes in minutes. And the rehearsal room. Voice mode is perfect for brainstorming on a walk, or rehearsing a pitch with the AI playing a skeptical buyer. Always end a voice session with a written summary.

3:55 Connectors and agents

Finally, the assembly bench. Connectors let the assistant read your Drive, Notion, Gmail or Slack, so you can say, pull together last month's campaign reports and summarize what worked. Agents go further: ChatGPT Work, Claude's agentic mode, Copilot's agents and Notion's Custom Agents can carry out multi step tasks, like building a competitor comparison spreadsheet or compiling a launch checklist from files across your drive. Start read only, keep approval on for anything that publishes, sends or deletes, and check the finished work like any other AI output. Agents save the most time on gathering and assembling. They should never be the last pair of eyes before something goes public.

4:43 Example 1: the review themes

A simple example. A skincare brand exports three hundred product reviews, removes names and emails, and uploads the file. The prompt: find the top five praise themes and top five complaint themes, with two example quotes each, and count roughly how often each appears. In a couple of minutes, the brand learns that customers love the texture but are confused about when to apply it in their routine. That's two wins: a caption angle straight from customer language, and a how to use post that answers the confusion. Before this, nobody had time to read all three hundred reviews.

5:26 Example 2: Leena's launch week

Now a realistic launch week, with illustrative details. Leena is the only marketer at a fitness apparel brand in Dubai, launching new leggings. Monday: deep research on how three regional competitors position performance leggings, with a table of claims and prices. She spot checks five links. Tuesday: she uploads anonymized reviews of their old leggings and finds the top complaint was see through fabric in squats. Wednesday: in the brand project, she drafts the product page, three captions and a launch email, using the review language, finally, no see through squats. Every output ends with facts to verify. Thursday: she rehearses her pitch to a retail partner in voice mode. Friday: an assistant connected read only to the team drive compiles the launch checklist. Every public claim is checked against the supplier's spec sheet.

6:24 Watch me do it

Let me build one. I create a new project called Leggings brand. I paste the instruction template from the lesson text and adapt it: our audience is women who train four times a week, voice is direct, warm and a bit funny, never invent statistics, testimonials or performance claims, write check instead, and end every output with facts to verify. Now I upload the brand kit: a one page voice guide, five of our best posts, and the product sheet. First test in a plain chat: write a launch caption for our new leggings. Result: elevate your workout with our revolutionary leggings, plus an invented sweat wicking claim. Same request inside the project: three captions in our voice, fabric facts from the product sheet only, and facts to verify: confirm the size range. That's the difference.

7:23 Common mistakes

The common mistakes. Putting client files into a project in a personal account. Projects are convenient, but they don't change which data terms apply, so client work goes in a work approved account. Stale product sheets, which quietly produce last season's prices. Update the files when anything changes. Treating a deep research report as final without spot checking sources. And giving connectors or agents permission to publish or send on day one. Start read only, approve actions, and expand only when the outputs have earned your trust.

8:01 Recap and try this now

Let's recap. Set up one project per brand or client with instructions, a brand kit, market context and a do not list. Use deep research for cited competitor and audience reports, file analysis for reviews and campaign data, voice for brainstorming and rehearsal, and connectors and agents for gathering and assembling, read only first. Here's your try this now. Spend thirty minutes following the hands on steps in the lesson text: create your brand project, paste the instructions, upload the kit, and run one real caption request inside and outside the project. When you see the difference, you'll never go back to the blank chat box.