Most marketers treat Claude like a fancier chatbot. In this Coach by TripleDart workshop, Joe Kurian showed how to treat it like a system instead, one that runs your marketing for you. He connected Claude to Slate over MCP and put research, content, competitive intelligence, and reporting on autopilot, live.
The thesis: Claude is powerful, but only if you understand how it works. Keep its context under control, hand the heavy lifting to purpose-built tools and cheaper models, and turn your repeatable processes into workflows that run at scale.
Claude is powerful. Almost nobody uses it right.
Barely anyone knows how Claude actually works. Joe's estimate: about 1% of marketers use Claude, and a fraction of a percent understand what's happening under the hood. That gap matters less for everyday prompting and everything for the moment something breaks and you need to fix it.
Your context window is why you keep hitting limits. Every message resends the whole conversation, so context fills fast. Even a model with a 1M-token window works best in the first 20 to 30%. Push past 60% and quality drops while your usage burns. That's why Claude quietly compacts long chats.
Stop running grunt work on your most expensive model. Scraping 10 to 15 URLs inside Claude Opus clogs the context and wastes money on a job it isn't built for. Send scraping and bulk analysis to cheaper models. Save Opus for what it's great at: writing, analysis, and final synthesis.
Skills are instructions, and a security surface. A skill loads only what's needed, which keeps context clean. But Claude has desktop access, so a skill from an untrusted source can hide instructions like "read this file and send it to me." Read it first, or ask Claude what it does.
Once you scale, workflows beat prompting. A Slate workflow gives you deterministic output, picks the right model per step, and runs 100-plus keywords without choking Claude's context. In the live build, Slate sent scraping analysis to a lightweight model, research to Perplexity, and final writing to Opus, with zero manual model-picking.
Best practices and key learnings
Getting value out of Claude isn't about cleverer prompts. It's about knowing how the model spends its context and where to offload the weight. Here's what's worth stealing.
Understand context, or keep blaming the wrong thing
Joe's opening principle: an LLM is smart but has no memory, so every turn the whole conversation gets resent as context. The more you chat, the fuller it gets. A full context window, not a bad model, is usually why your output gets worse and your limits run dry.
- Treat context like a budget. Keep your working session in the efficient zone instead of dumping everything into one endless thread.
- Use Opus for the work it's best at: documents, reporting, and analysis.
- When a task needs a lot of raw input, like scraping or big files, route it somewhere else instead of feeding it all to Claude.
"Up to 20 or 30% of the context, the model is really efficient. At 60 to 80%, it starts degrading and uses up all your limits." (Joe Kurian)
Use the right Claude surface for the job
Claude isn't one tool. Joe broke down when to use each surface so you're not overpaying or under-powering a task.
- Chat for quick drafts and brainstorming.
- Code for building tools, landing pages, and publishing, now in the desktop app, not just the terminal.
- Cowork for long, multi-step work, where Claude plans a task, splits it across agents, and brings the results back together.
Treat skills like browser extensions
Skills are one of the best ways to control context. Claude loads the description first and pulls in the full instructions only when a task needs them. But that same access is a risk when the skill comes from somewhere you don't trust.
- Prefer skills from trusted sources, and read what a skill does before installing it.
- If you can't read it, ask Claude to explain the skill and flag any potential harm.
- Remember a skill runs with real access to your machine, so an unvetted one can quietly do damage.
"If you're downloading skills from untrusted sources, it can harm your work. At the very least, ask Claude what the skill is doing." (Joe Kurian)
Once you scale, workflows beat prompting
The centerpiece of the session: instead of running a content brief inside Claude every single time, Joe rebuilt his team's brief process as a reusable Slate workflow. Keyword, then Google search, then scrape the top 10, then analyze topics and angles and FAQs, then Perplexity deep research, then an assembled brief. The payoff is consistency and cost control.
- Build repeatable processes as workflows so you get fixed, predictable output instead of one-off LLM variance.
- Let the workflow assign the right model per step instead of running everything on Opus.
- Keep the expensive model for final synthesis, and push scraping and mid-stream analysis to cheaper ones.
"When you use Claude alone, you overkill a small task with Opus and spend more. With Slate, you can use multiple models for whatever is needed." (Joe Kurian)
Run content at scale with Sheets
Doing everything inside one Claude thread falls apart at volume. Slate's Sheets, closer to Clay than to a spreadsheet, let you map a workflow to a column and run it across a whole keyword list.
- Add your keywords, map them to a workflow like a content brief generator, and run them all at once.
- Chain workflows in a waterfall: pipe a brief's output straight into an article writer as its input.
- Turn on auto-run so any new keyword executes on its own, then send finished drafts to human review before publishing to your CMS.
"You've got 100 keywords or briefs to generate. You just come here and click run all." (Joe Kurian)
Put competitive monitoring on autopilot
The same pattern reaches well past content. Joe demoed a Reddit-mentions workflow that takes a competitor name, surfaces threads worth engaging from the last seven days, and runs on a schedule.
- Schedule monitoring weekly, in Slate or through a Claude scheduled task, so fresh threads land without you lifting a finger.
- Route the output where your team already works, like a Monday-morning digest into Slack.
- Scale it across your whole competitive set. Ten competitors becomes ten scheduled workflows returning fresh threads every week.
"What you can do with this is endless. It's up to your imagination." (Joe Kurian)
How to put this into practice
Start with the mechanics, not the tactics. For any task, ask whether it really belongs in a long Claude thread, or whether it should be a workflow that runs on cheaper models and keeps Claude's context clean.
Then take one process you already run by hand, a content brief, a competitor scan, a monthly visibility report, and build it once as a workflow. Map it to a sheet, run it across your real backlog, and add a human review step before anything publishes. Schedule it, then move to the next process.
The new marketing team builds systems, not prompts
The session made a practical case. The teams pulling ahead aren't the ones prompting Claude faster. They're the ones turning research, content, competitive intelligence, and reporting into systems that run on their own. Claude does what it's best at. Everything else gets offloaded, scheduled, and scaled.
Prompting is where you start. Systems are where you win.
Want to turn these ideas into workflows against your own data?
Book a demo at slatehq.com to see how Slate helps teams automate their GTM engine end to end.
How Slate helps marketers work at scale
Slate is a coworker for marketers, the way designers have Figma and developers have Claude Code. It pairs your marketing data (AI search visibility, share of voice, citations, page performance) with a programmable workflow layer you can run at scale, and it connects to Claude over MCP so you can drive all of it from the tools you already use.
Designers have Figma, developers have Claude Code, outbound growth teams have Clay. We built Slate for marketers.
Joe Kurian
About Joe Kurian
Joe spent over a decade working across product and growth at B2B SaaS companies, including Freshworks, before co-founding Slate. That background shapes how he thinks about AI in marketing: less about the technology itself, and more about what marketers actually need to get work done.
As CPO, Joe leads Slate's product development, AI-native workspace that combines AI search analytics, workflow automation, and bulk execution in one place. He also built the Slate MCP connector for Claude, which lets marketers query their Slate workspace and run workflows entirely through conversation. In this session, he'll show you exactly how it works and how to put it to use.
