Head of Product Marketing at SaaS startup on automating product marketing with Claude Cowork

We spoke with a product marketer at a retail-focused startup who uses Claude Code, Cowork, Skills, and Design as the backbone of her solo marketing operation.

The conversation covers how she builds and debugs recurring agent workflows for competitive intelligence, hiring signals, and industry updates, and where human review remains non-negotiable before anything reaches customers or leadership.

Key points via Sacra AI:

Which tools do you use regularly, and which have you tried between Claude Cowork and OpenAI's Codex?

I use everything in the Claude ecosystem—Claude Code, Claude Cowork, Claude Skills, Claude Design. I use Codex occasionally, but not too much.

What does a typical daily Cowork task look like for you—the last thing you handed off to it, and what came back?

I'm in product marketing, and I do that for my whole company—I work at a startup, so the context is probably a little easier to explain. I do a lot of research and analysis. I have Cowork do monthly industry updates on the search world. I have it find job openings in search at big retail companies that are our prospect accounts. I also use it for scheduling, like if I need to meet with a partner. And I use it for competitive intelligence—I can point it at a competitor's website and have Cowork do an analysis of what's different, kind of like a SWOT analysis.

It's all marketing-specific work. I do a lot of automated updates and send them directly into Slack—research, competitive intelligence, industry updates, all pushed straight there.

Could you walk me through one of those recurring workflows end to end—maybe the monthly search industry update? What exactly do you ask Cowork to do, what sources does it use, what does the output look like, and how does it get into Slack?

Slack has been really good at working with Claude. It's basically a Slack API for Claude—Claude Code within Slack. You can integrate Claude into a Slack workspace and converse with it there, including delegating tasks to the team, like following up on scheduling meetings for an event. It can also analyze recent messages for context. You can create a Claude Code session and post progress and completion updates in the thread. It also supports a GitHub account, and you can use Claude Code on the web, which is super helpful.

Talking about the end-to-end workflow: there are skills, connectors, and plugins. Skills are the reusable instructions for how Claude will think and format a task. I have a skill that defines how I want the Cowork output to look—I want bullets, no em dashes, a recap of what the announcement is, what it means for retailers, and how Optiversal helps. That's part of the skill. Then I connect data sources, which can include Google Drive, Reddit—which is really popular right now—YouTube, and some search newsletters I pull from. Then there's the plugin, which is how the connector is built.

What I like about Claude is that it will follow up with instructions about the workflow—tools, edge cases, existing documents it should pull from to map relevant parts of the plugin.

One thing that has been a struggle in this workflow is that Cowork has no memory between sessions. Every conversation starts fresh, which means if you don't have the right setup, it doesn't remember who you are, what you do, how you like to get things done, or how you like to format things. You need a context file—an MD file with the role, the brand, the voice, the goals—in the project folder every session so it already starts with that context. You can go to the Cowork tab, type `/setup cowork`, and it walks you through setting up the context and the plugins.

Can you walk me through the last time that workflow actually ran? From the moment you kicked off the industry update, what did Cowork pull in, what did it produce, and did you have to edit or correct anything before sending it to Slack?

It actually knew a lot of the external sources to pull in. It's worth thinking about external versus internal use cases—if I'm asking for industry updates, I can say, "What did Forrester say, what did Bain say, what did these newsletters say, what are these experts saying on LinkedIn?" I did include some specific links to research manually, but it did a lot on its own in terms of sources. I didn't have to change much. I did adjust some of the output formatting, but I can chat with Claude within Slack, so those updates are easy to make. I just had to make sure I updated the skill as well so the formatting stayed consistent.

It actually runs weekly—it's a weekly update. Another interesting one is the job openings pull, which finds who's hiring for search so we can see which companies are making search a priority, or whether someone new is coming in as head of e-commerce. That's a good one too, because job openings are posted across so many sources—LinkedIn and countless other job posting sites. It all integrates through the Slack API and pulls into the weekly update.

I changed a little bit of the format along the way, and there were some buzzwords I didn't want it to use. I also had to provide a lot of context on what Optiversal does so it knew how to connect product relevance to the trends it was pulling. That was pretty easy to do by connecting some internal documents. We also had a Google NotebookLM with information about our different products that it could pull from.

Can you walk me through the job-openings workflow end to end—what prompt or recurring setup you use, which companies or roles it tracks, what Cowork returns, and how your team actually uses those hiring signals?

The prompt is something like: what retail companies—since we're retail-focused—are hiring for search roles at the director level and above, within the last month? I included specific titles. I also prompt it to include a link to the job opening even if it's closed, as proof that it was a real posting and to guard against hallucination. The sales team can use that link to say, "Hey, you had this job opening—did you hire? How can we help get that person onboarded?" Our customer success team also finds it really useful. If we have an existing customer who's hiring a new head of search, that's really important to know.

The way the team uses it: CS can say, "I know you're hiring—can we help them get onboarded? Do we have recommendations?" And sales, while prospecting, can say, "Obviously search is a big priority for you. I see you're hiring for this role. How can we help get a good foundation in place before that person starts?"

When Cowork returns that hiring-signal list, what does the actual artifact look like? Is it just a Slack post, a table, a doc, something that goes into Salesforce or HubSpot—and who reviews it before sales or CS acts on it?

No one reviews it—no human. I will see it as it's posted in Slack and do a quick double-check on the links just to be sure. It doesn't include contact information for the new role, and it's not put into a CRM—it's just in Slack for sales to follow up with immediately.

Cowork can also populate an Excel spreadsheet. What I've done is ask it to put the list of companies, the role they're hiring for, and the link into a spreadsheet so sales can track which ones they've actually contacted. That way it doesn't just get lost in a Slack channel.

The nice thing about Cowork within Slack is it can also do progress and status updates. I don't have to pay for something like Asana. At a startup with a small team—three salespeople, three in CS, and just me in marketing—it's easier to track this way. But that's something worth considering for larger teams.

One of the things I really like about Cowork is all the integrations—easy preset APIs. For example, Cowork can do work in Canva. If I need a social image and I describe what I want to Claude, Cowork can actually help create the design in Canva.

Can you walk me through the Canva example—the last time you used Cowork to create a marketing asset? What did you give it, what did it produce, and how much manual cleanup did you have to do?

I use it a lot for social posts. Connecting to Canva is easy—if you go to settings in Claude, you can find the Canva connector, hit connect, and it opens a browser login. You don't need MCP or Node.js or anything like that. In Claude desktop, you can also click "Customize," select connectors, and click the plus icon to add Canva that way.

Claude Design is also doing a really good job creating images, slides, and presentations, but you can't edit the output. That's the only thing Canva is better at right now. The Canva connector lets Claude create new presentations, resize designs, autofill templates, and search and summarize content from across the Canva workspace—which is helpful if you have a workspace dedicated to your company's work. And it all responds to natural language prompts. You're not uploading files or manually transferring ideas. You can brainstorm with Claude, and that turns into a Canva design.

I've done one-off social posts and social carousels—for example, "Four steps to prepare for holiday 2026," since we're focused on retailers. You can brief Claude on the campaign, have it generate copy, and spin up matching Canva designs from a branded template using your tone, voice, and brand guidelines. You can also resize the Canva design into different social media formats—if you're posting on LinkedIn and also want it for Instagram—and export everything as PNGs.

You can also have Claude search your Canva workspace for existing assets to reuse rather than starting from scratch. It really does help from brainstorming through creation and editing, because it's conversational and that conversation turns into design.

That said, I still use a designer to tweak. I wouldn't say I use this a hundred percent autonomously. A lot of the editing isn't there yet, and some of the images end up really crowded with text or look complicated. But it's a really good starting point.

When you're deciding where to do a marketing asset—Cowork, regular Claude chat, Claude Design, Canva, or even ChatGPT—what makes you pick one over the other? Is it editability, integrations, brand context, or just habit?

Maybe all of the above. The biggest issue with AI design right now is editability. Pitch is actually doing a decent job—you can conversationally say you want a slide deck that looks a certain way, it uses your brand guidelines, and it's editable. Claude Design isn't there yet. Gamma isn't there yet. ChatGPT can make a decent image, but it's really hard to edit individual pieces. Even if you say "move this text" or "remove this part of the image," you have to do it one at a time. You can't give it ten edits and have it make them all in one round the way an actual human designer could.

Connections definitely help. I'm just used to working in Canva, and because Claude has a Canva connector, that's super helpful—no MCP or anything like that needed. I'm also building my whole ecosystem in Claude with skills, context, and everything else. Partly it's habit, but it's also that I'm building my personal foundation in Claude, which makes everything easier. But AI is still really limited by editability and brand accuracy. You can add brand guidelines, but a lot of what's generated—including by Claude—doesn't get the typography right. I end up having to take it to a designer. It's good for inspiration and a first draft, but getting to a final asset takes a lot of back and forth.

Can you walk me through a recent presentation or deck workflow where you used Claude, Cowork, Pitch, or Canva? What did the agent do versus what you or a designer still had to do manually?

With Claude Design—it's also really good for ebooks, but same issue. I can get a draft of an ebook, get some ideas for visuals and spacing, and see what it looks like across pages. But I can't say, "You're missing a word here, fix it," or "The spacing is weird," or "This should be bulleted." It's very much inspirational.

Pitch is getting better. You can ask it to make a deck on a certain topic and it will apply your brand guidelines, and at least that's editable. It's a bit easier for decks that don't need too much design work. But I'm still going to a designer for custom images or for describing exactly what I want an image to look like.

The other big problem, besides editability, is product images. If I have a screenshot from within the product and need to tweak it or anonymize it—that's a huge one. Anonymizing a screenshot from a customer so it doesn't include their name or specific revenue numbers is really hard to do without a designer. So I'm still having to edit. There's no single tool out there, including Cowork, that gets me to a hundred percent.

Can you walk me through the last time that broke down? What screenshot or product image did you give it, what did you ask it to anonymize or change, and where did the workflow stop being useful?

If I say "anonymize," it doesn't even know what that means. You have to have a human eye to catch things—I don't want revenue numbers visible, or anywhere the logo appears, or a brand URL that would give away who the customer is. It won't find the brand name embedded in a URL. So there's just a lot of human review required.

When Cowork or Claude gives you work back—whether it's a Slack update, spreadsheet, competitive analysis, or design draft—what's your review process? What do you check most carefully before you trust it or share it?

It depends on the asset. If it's just a bulleted list internal to Slack, I can always delete it. But external stuff I still don't trust. I wouldn't have Cowork autonomously post to our LinkedIn page—the words still sound like AI. There's something really important about that human voice. Almost anything going externally, I won't just let an agent post autonomously. Same reason I wouldn't let an agent shop for me—I wouldn't hand an agent my credit card and say "buy this when it's back in stock." There's almost nothing I do on a set-it-and-forget-it basis, other than bulleted updates within Slack.

Can you give me one specific example where the output looked plausible at first, but you caught something that would have been risky or embarrassing if it had gone out?

Visuals—maybe my logo wasn't right. But a more substantive example: I asked Claude to build me a social calendar for the summer with some topics and product context. Some of the posts, if I had let it post autonomously to LinkedIn, had messaging conflicts. It was saying things like "you shouldn't rely on AI to build all this content," while our actual product uses AI to build content on behalf of our customers. There was a contradiction between what the social posts were saying and what our product does. I had to prompt it and ask why it was saying that and where it got that information—and the source turned out to be some random Reddit post.

So almost everything external has to be checked by a human, whether it's going to customers or even just in slides.

When that happens, how do you repair it? Do you update the skill, add more internal source-of-truth docs, change the prompt to rank internal docs above web sources, or something else?

It's a little bit of all of the above. I do have to press it on where it's pulling data from and why it's saying something. Part of the problem is that AI likes to agree with you—"Oh, of course I shouldn't have used that source," or "Of course I didn't mean that, I meant this." Even with tone—even in skills, I'll say don't use em dashes, and sometimes it still uses em dashes.

It's a combination of how I'm prompting and sometimes needing to update the skill with very specific instructions: do this, don't do this, and here's a specific example. When you're updating a Cowork flow, there are a lot of moving parts to fill out, and if you haven't done this manually before, it's really tricky to know where to fix it. You have to know where it's breaking in order to know what to fix, and that can be hard to diagnose.

Can you walk me through one specific Cowork flow you've had to debug like that? What was going wrong, and what did you actually change to get it working better?

The competitive intelligence workflow was like that. It was actually pulling landing pages from a competitor—pages specifically about why they're better than us—and using those as key cited sources. Of course the competitor is going to say they're better in certain ways, even where we're actually stronger. Our own product documentation would have debunked those claims.

So I had to prioritize certain documents—look here first. I added a step in the workflow to double-check product proof: if you're writing something, make sure you check these documents or this workspace first. It came down to either adding or removing a step, or fixing a prompt within a step.

The thing I'd love is if Cowork could, within a workflow, identify where it has questions or where it's getting confused. Think of it like Shutterfly—when you upload an image to make a product, it shows a little exclamation point and says "this image is blurry" or "this font is too big" before you make the purchase. A similar warning system in a workflow would be really helpful—something like, "I'm not sure about this source," or "I'm pulling from a competitor site—make sure this source is valid." Flagging the potential problem so you can scroll over it and see what the concern is and what the potential fix might be.

What parts of Optiversal—your product, customers, or messaging—do you find yourself repeatedly re-explaining to Cowork, even with skills and context files set up?

A lot of product details—what our tool does and doesn't do. I'm consistently having to update what the key customer pain points are that we're trying to solve. That's important because it changes over time. So it's always evolving, and if I don't review the output, it's worth asking: what is it doing wrong, and why?

How do you maintain that today? Is there one doc or folder Cowork should treat as canonical, or is context spread across Drive, NotebookLM, Slack, your own head, and older collateral?

It's everywhere, especially since I switched from ChatGPT to Claude. It's in Google NotebookLM, Perplexity, Google Drive—so many places.

What I'd love is if someone onboarding to Claude for their company had a checklist: here are all the documents you need to have a really good context—a really good brain. And then proactive reminders: "This document is over a month old—can you check it and make sure it's the right version, or does it need to be updated?" Being proactive about maintaining the context would be super helpful, because right now everything lives in a million different places—our website, our CRM, Gong calls—and you're connecting a million sources. I never quite know if I've connected all the right ones, or if there's a source type that would be really helpful for a certain kind of deliverable or workflow.

There was also an AI company that had prebuilt workflow templates, where you could build an agent by starting from a template, customizing it, and being walked through the setup. That was really helpful.

When you mentioned that AI company had prebuilt workflow templates—for an ideal Cowork template for something like competitive intelligence or weekly industry updates, what would it ask you for upfront, and what best-practice steps would it already include?

It already had all the workflow steps set up. If you've never done a workflow manually, I don't think you should be trying to build an agent for it cold. But with templates, it was like: if I want an agent to write my monthly customer newsletter, here are the steps—look at these resources, read the product roadmap, read the release notes, here's the design format. Step by step, it was already built out. Rather than having to build the steps, then test, then build more steps, you were mostly just plugging in the sources. You were connecting rather than constructing.

When you think about Cowork versus regular Claude chat, what makes Cowork feel more like delegating to a worker with a repeatable process rather than just prompting a chatbot?

Cowork just has so much more to it. With ChatGPT, you can build custom GPTs and agents of a sort, but it doesn't have the same cohesive ecosystem. Claude has Claude Code, Cowork, and Skills, and you understand how they all interact. Claude does a better job of building an ecosystem where everything builds on each other—you need Skills to do Cowork, so it all connects. And the context is already there because that's where you're building your whole foundation.

Can you compare that to Codex? Since you've tried it a bit, where does Codex actually fit for you today, and why hasn't it become part of the same daily workflow?

Claude Code is more interactive—it's local to my computer, very collaborative. Codex CLI is more task-execution focused. It runs in a cloud sandbox environment versus Claude Code, which runs in your local terminal and local file system, so it has deeper access to your real-world files. Codex may be stronger at multistep automation tasks, back-end logic, or benchmark-heavy coding performance—"build me this end-to-end task." Claude Code is better at complex refactors across many files, debugging with reasoning and explanation, and it keeps the code more readable and human-style. Claude Code also supports really large context windows, which helps when your repo is big or you need deep cross-file reasoning.

Codex chunks tasks more aggressively and relies on structured tool execution rather than full-repo reasoning at once. The feel is different: Claude Code is like, "Let me inspect this file—here's what I think is wrong, want me to fix it?" Codex is more like, "Task received. I will implement, test, iterate." A lot of people choose Codex because it's cheaper depending on your ChatGPT plan, but I still prefer Claude Code.

Since your main work is product marketing, not engineering, can you walk me through the last time you used Claude Code or Codex where the output wasn't really code but something broader—like an internal tool, analysis, dashboard, workflow, or report?

I've done some funnel reporting. I've also turned things into interactive landing pages, though Lovable is the best by far for that. The reporting is trickier to set up. Product marketing is really hard to measure, so I actually asked Claude for recommendations on what to track for product marketing specifically, and then built a dashboard using those recommendations.

Can you walk me through that product marketing dashboard end to end? What data sources did you give it, what metrics did it recommend, what did it actually build, and how are you using it now?

The metrics I was looking at: revenue, customer acquisition cost, conversion rate, churn rate, opportunity status—stage one through close—how quickly prospects were moving through stages, and where prospects came in from: an event, social, a connection. All of that lives in our CRM. We're a small company, so we don't have as many sources as others, but it was connected to HubSpot. Salesforce could work too.

What did it actually produce at the end—a spreadsheet, a coded dashboard, a HubSpot report setup, or something else? And how much did you have to adjust before it was useful?

I did have to adjust it quite a bit. It did a really good job on an Excel spreadsheet, but that wasn't real-time data—it was a specific point in time. I couldn't go in and compare today versus yesterday, or year over year. It was static. It can create a HubSpot report internally, but I really wanted the visual so I could then have conversations with Claude—like, "Why do you think our churn is happening?"

I also linked it to Gong to pull in where customers had objections or expressed concerns—competitor mentions, dissatisfaction signals, things like that—and then ask, "Why do you think customer churn is increasing?" It could pull from Gong calls for accounts that were churn risks and surface what was important from those calls: the warning signs, what they were complaining about.

A simple version of this prompt would be something like: "Create a single-file HTML dashboard with D3.js showing sales data by region, by industry, and by stage. Include a daily revenue timeline, a product category donut chart, a top-ten customer table, dark mode, animated counters." Then Claude Code can write Python to connect to your database, list all tables, get the full schema, and run sample queries—without you having to write that code yourself.

When you had Claude connect CRM and Gong context, what was the hardest part: getting access to the data, shaping the right metrics, making the dashboard update live, or trusting the analysis it gave you about churn and objections?

All of the above. It's scary to trust this, especially when you're presenting it to your CEO. I did manually check a lot of it to make sure it was reporting correctly. But it was actually pretty good at explaining why customers might be churning. Salespeople typically say "it's a competitor" or "it's price" with no real context. Claude was better at pulling out context and warning signs—why someone might churn or why they actually did. That was super helpful. But anytime I'm doing monthly reporting to my CEO, I have to be really sure the numbers are correct.

Walk me through your review process for that dashboard or monthly report. What do you manually verify first, and what parts of the agent's analysis are you more willing to trust?

The hard numbers—revenue, things that come directly from the CRM—I verify. But the intent signaling is where I'm more willing to lean on it. Things like: if a customer is at risk of churning and someone hasn't updated the CRM, can it send a Slack notification ahead of time? "Hey, this customer is struggling with this"—helping CS be more proactive with warning signs. And on the flip side, when customers are really happy, pointing out quotes I can bring to CS and say, "I want to use this in marketing."

From a dashboard perspective, the hard-coded numbers I can cross-check in the CRM. I can manually verify revenue with a quick gut check to make sure it's still reporting correctly. If most of the numbers look right, I can generally assume the dashboard is right. But if my CEO questions something—like how quickly we're moving from stage one to close—I have to be able to justify that and know where the data came from. And if there's a drastic change, I can ask Claude why that might have happened, especially with additional context, but ultimately it's on me to explain it to my CEO.

If you imagine the ideal version of this dashboard and reporting workflow, what would make you comfortable relying on it more—source citations, data lineage, anomaly explanations, approval steps, or Slack alerts when something changes?

All of those. If there's a drastic drop in revenue, or we're eighty percent behind budget for the quarter on new business—proactive alerts. Approvals: "Does this look good? Yes? Okay, publish." All of that would be super helpful.

How much of your Cowork usage today is just you as a power user versus something the broader sales, CS, or leadership team is actually adopting with you?

I build everything. I'm the one executing, and they take in the data and information I'm sharing. In terms of adoption, they're consuming the outputs of what I'm doing in Cowork. It's really just me as a power user—they're all marketing agents at this point.

What I'd love to do—and haven't tried yet—is let sales put in a prospect's company name, their name, their title, and get back a business plan: how to reach out, what email campaigns to run, what assets to send. Some kind of business plan automation for sales outreach would be really valuable, but I just haven't built it yet.

If you were going to build that sales outreach workflow, what would the ideal input and output look like? What would a rep enter, what sources should Cowork check, and what would it hand back that's actually usable?

That's partly why I haven't built it yet. The input would be company name, person's name, and title. Sources could include recent news: how long have they been in this role, have they achieved anything notable, what do their press releases look like, anything new on their website, company announcements, earnings calls—earnings calls are huge, especially if you can analyze the audio and pull out things like whether they met their goals and what their strategic priorities are this quarter. Reddit signals. New product launches. New initiatives.

So very minimal input from the sales team, a lot of sources being checked, and then the output type as an additional input—does the rep want a full business plan, a series of three outreach emails, a designed one-pager on a specific product? There are really interesting tools for custom sales assets where you enter a title, pain points, and product details, and it builds a custom asset. Something like that would be super helpful.

What would need to change for sales reps to actually use that themselves instead of relying on you? Is it trust in the output, ease of setup, integrations with CRM and email, or knowing what to ask for?

If you make the inputs easy—basically "enter company, person, title, choose output"—there's nothing preventing them from using it. I just haven't built it yet.

What would make you comfortable letting them use it without you reviewing every output first? Would you need approvals, citations, locked messaging guidelines, or limits on what it can send externally?

A human needs to review anything before it goes to a customer. I don't trust AI SDRs or BDRs because they send the same message to a hundred people with no real personalization—I've received so many of those myself. But beyond that, there need to be citations, alerts if something feels off-brand or off-message, and metrics. I need to know what's working. If I can say this agent is helping the team move through stages faster, shorten the sales cycle, or influence a certain amount of revenue—if there's data behind it—adoption would be super easy to justify, especially in a bigger company. A dashboard showing when sales uses this agent for this purpose and here's what that results in would make the business case essentially build itself.

Walk me through how you'd justify paying for Cowork internally today. What's the clearest business value you'd point to—time saved, better sales signals, more output from marketing, or something else?

All three of those, honestly. But I pay for it myself, which isn't ideal. Our company pays for Gemini, but I don't like Gemini, so I pay for Claude myself. To actually justify it, you'd need all three.

If you had to make the case to your company to switch or add Claude and Cowork, what proof would you bring? Would you show hours saved, specific sales opportunities surfaced, assets produced, or some kind of before-and-after workflow?

Maybe assets produced. I think a lot of times it comes down to security or internal tools—I know I have a lot of friends in...

This transcript is for information purposes only and does not constitute advice of any type or trade recommendation and should not form the basis of any investment decision. Sacra accepts no liability for the transcript or for any errors, omissions or inaccuracies in respect of it. The views of the experts expressed in the transcript are those of the experts and they are not endorsed by, nor do they represent the opinion of Sacra. Sacra reserves all copyright, intellectual property rights in the transcript. Any modification, copying, displaying, distributing, transmitting, publishing, licensing, creating derivative works from, or selling any transcript is strictly prohibited.