Matt Swulinski spends $4,000 to $4,500 a month on Claude. He's Wispr Flow's head of growth, and outside the engineering team, he's the company's top AI user.
That spend isn't going toward polished copy or generic automation. It runs what he calls the Wispr Flow Marketing OS – a Claude system that handles everything from 100 monthly newsletter sponsorships to investor updates.
It got good enough that Wispr paused hiring for two marketing roles, because the existing team could absorb the work. He built it to survive doing marketing solo until December. Now he's rolling it out to a team of nearly 20.
In our in-depth conversation, we discuss:
- Why spending $4,500 a month on Claude is the cheapest hire he's ever made
- The "I wanna build an operating system" prompt that makes you intermediate by the end of a single session
- Why he stopped fixing bad copy himself, and how that one habit made every future draft sharper
- The quiet mistake that breaks almost every AI system, and the single-file fix
- Why he'd take an intermediate systems thinker over the best Meta buyer in the world
The skill that started it all
Newsletter sponsorships are full of admin: negotiate, sign a contract, write copy to spec, drop a tracking link, analyze, repeat. "Now multiply that times around 100 newsletters that we run placements on any month," Matt said. "Without AI, it's essentially impossible if you're running anything else."
He built a master skill in Claude Code that calls other skills as it goes. He drags a contract PDF into the terminal, and Claude checks the CPM, researches the audience, and pushes back if the price or fit looks weak. If the deal moves forward, it writes copy in Wispr's voice and builds tracking links – all of it drawing on performance data from every newsletter Wispr has ever run. The skill saves time and makes every placement smarter than the last.
How the OS is built
Matt runs six to ten Claude Code sessions at once, each focused on a different channel. The whole system lives in a Git-controlled folder shared across the team, so anyone can pick up where he left off.
Two skills do the heavy lifting: session-start and session-end. Session-end logs what got done and clears the context to keep costs down. Session-start pulls those summaries back in and surfaces unfinished high-priority work. With 100-plus open to-dos at any time, it's the only way he keeps things from slipping.
The mistake he warns against is hard-coding numbers into skills. ARR, team rosters, Slack IDs – all of it goes stale the moment you share it. Keep variable data in one central file and have the skills pull from it, so you only ever update in one place.
How to hire for this
Two years ago Matt would have hired the best in the world at Meta or Google. Now he wants systems thinkers. "I would much rather the person be intermediate to advanced, but be a systems thinker who can understand all the moving pieces and layer in AI to 10X themselves," he said.
The test isn't whether someone can use a chat window. It's whether they can deconstruct a workflow, find the bottlenecks, and map the inputs and outputs. If they can, they can build a system. If they can't, they'll just use AI to send emails faster.
Where to start
Matt built the original OS by asking Claude to help him build one. "Start there and you'll already be an intermediate at the end of that session, because you'll have something that is unique to you."
His advice: start small, build one or two workflows around your own bottlenecks, and don't copy someone else's system. And assume everything you build gets shared eventually – because that's what makes the whole team more powerful.
