Last week, Torsten Sandor ran a live workshop on Substack. The topic: how to actually work with AI agents — not as a chatbot upgrade, but as a co-worker that handles research, analysis, reporting, and execution while you focus on the decisions that matter.
We called this workshop part of our new Teamless initiative. The idea is simple: the teams that figure out agents now will need fewer people to do more work. That cuts both ways — as a leader managing headcount, and as a professional managing your own career.
The recording is now on YouTube. Watch it here.
If you’d rather read the summary first, here are 10 things you’ll learn from watching it.
1. Agents do work. Chatbots just respond. A chatbot answers questions. An agent opens your browser, reads competitor websites, writes a Word document, saves it to a folder, and moves to the next task — without you touching a thing.
2. Claude Cowork sits between Chat and Code. It’s Claude’s desktop agent. More capable than the web chat. Much friendlier than Claude Code, which is built for developers. If you’re a non-technical marketer, this is your entry point into agentic work.
3. Projects are just folders — and that’s a good thing. Cowork organizes work around projects. A project can be a campaign, a quarter, or your whole brand. Everything Claude learns and builds stays inside that project context and compounds over time.
4. One prompt. Five competitor websites. One Word doc. Torsten asked Claude to research five baby-wearing brands, extract key messages, voice, tone, positioning, and competitive gaps, and deliver it as a formatted document. It would have taken a junior marketer three days. It took Claude about ten minutes.
5. Agents get smarter the more you give them. After the competitive research, Claude built a tone-of-voice guide and a key messages document from scratch — based solely on what he had already created for the project. No new uploads. No re-briefing. That’s the compound knowledge effect.
6. You can build an email drip campaign in a single prompt. Using nothing but the brand context it had already built, Claude wrote a five-email welcome sequence with subject lines, preview text, and full copy. If you have HubSpot connected, it can upload those emails directly.
7. Agents can connect to your existing tools. HubSpot, Supermetrics, CMSs, analytics platforms — Claude Cowork has connectors for most of the stack you’re already running. Your agent doesn’t just write about your data. It can work with it directly.
8. Ad performance analysis in 20 minutes. No Excel formulas. Torsten uploaded a raw Meta Ads export. Claude wrote Python behind the scenes, ran cohort analysis across age groups and targeting approaches, and returned a formatted Excel sheet with color-coded analysis and a clear narrative on what was driving cost per conversion.
9. Skills turn one-off prompts into reusable workflows. Once Torsten had the ad analysis working, he saved it as a skill called “meta-ad-analysis.” Next time he runs the raw export, a single slash command will run the entire analysis automatically. Skills are repeatable work with zero repetition.
10. Scheduled tasks mean the work runs without you. Claude Cowork supports scheduled tasks. Torsten set the ad analysis to run every weekday at 4 pm, the interactive dashboard to generate at 4:30, and the report to email his manager at 4:40. The whole reporting chain runs automatically while his computer stays awake.
If this is a conversation you want to have for your marketing team — a hands-on workshop on building agentic workflows for real marketing use cases — reply to this email. We’re exploring customized sessions for teams that want to move from AI chat to AI execution.
