Jan Červinka

CTO and Managing Director @PurpleTeamLife · Building fintech infrastructure for the AI era

[Jan Červinka @JanCervinka]

When I first heard people at Every talk about their "Plus Ones" — OpenClaw agents living as coworkers inside Slack — I was intrigued, but also knew it was too much for us and too fiddly to set up well.

But the form factor of collaborative AI — an agent as an actual coworker, living where your human coworkers do — is really powerful.

Now Anthropic launched Claude Tag. This one seems promising, but also worth considering carefully. What permissions do you give it? What tools and skills?

Let's figure that out. I'll report back once we have some hands-on experience.

Share yours, if you've given it a go.

[Erik Voorhees @ErikVoorhees]

No, you don't get it.

He does not have $1 trillion sitting in cash, it is 99% stock in his companies.

To make that wealth liquid would mean selling all that stock which would swiftly destroy *both* the companies (Tesla, SpaceX, others) and the wealth. If he sold it all, he'd end up with maybe $100b max, several hundred thousand people would be out of work, the companies ruined and many of their suppliers also ruined.

Okay, but now Elon has $100b in cash, and can "solve the world's problems".

$100b divided by the world's 8 billion people is $12

If you were in charge, several of the most innovative industrial companies in the world would be destroyed, hundreds of thousands out of work, and space would again close to human civilization for another generation.

But everyone on earth could have one nice meal and you could revel in your altruism.

[Jan Červinka @JanCervinka]

"I apologize for the long letter. I didn't have time to write a short one."

That's been a joke for 400 years. AI made it the default.

A detailed review used to take real effort. Today you generate one in 30 seconds. Every section covered, every gap flagged, every imprecision raised. So that's what we send.

The work didn't disappear. It moved. The person on the other side has to read several pages, figure out which of the flagged points actually matter, and reply. Often with their own AI-assisted thorough response. And we're off.

The fix isn't to stop using AI — it's making us better at a lot of things. It's to take 30 seconds before hitting send and ask: what actually matters here?

I started seeing this at Purple. So I addressed it.

[Jan Červinka @JanCervinka]

@michalpur Já jsem nikdy o Ferrari nestál. Ale tohle by se mi líbilo. Takže souhlasím. Otázka je, jestli to je pro Ferrari dobře, protože já si ho stejně kupovat nebudu 😃

[Jan Červinka @JanCervinka]

For 15 years, we built fintech in pieces. Trading. Payments. Technology. Ventures. Built from the inside out since 2011. But we've never attempted to show the full picture. Until now. (with a bonus in the comments below)

https://t.co/gagg1TsPQZ is that picture. Not a new company. Not a rebranding. Just an overview of what the Purple ecosystem is all about.

But there is more. We are also creating a way for anyone to become an architect of their own financial product. To get a glimpse of that, check the link in the comments 👇

[Jan Červinka @JanCervinka]

Juggling multiple ongoing agents is hard mental work. Anthropic just released Agent View to fight that. But it's not enough. I still need to review the work they produce, keep track of the tasks and goals, and make sure I don't forget something in the process.

You can solve this with Linear, but that's built for collaborative development. Not personal knowledge work. So I put together a private alternative.

Obsidian Base + the Kanban view plugin + a `day-operator` skill in Claude Code that I run on a `/loop`. It picks up new tasks, writes a plan, and waits. I review the plan in Obsidian, leave inline comments, send it back. Next iteration the agent replans. When I'm happy, I move the card to "ready" and the agent executes. Every task has a changelog so I can scan what happened.

I also told the loop to watch my Monologue voice notes and starred Gmail items and create tasks from them. Now I can talk an idea into my phone or star an email and the agent picks it up.

OpenClaw never worked for me. Too many things to break. Too little focus on deep work that you can't do well through a Telegram message. This is a much simpler and focused alternative from existing tools that basically needs just a custom skill.

I can attend to things when I have time, not when an agent pings. And I don't fear closing the terminal anymore.

[Jan Červinka @JanCervinka]

Falling birth rates are a problem that does not seem to have a good solution, yet. The recent FT article (posting unlocked) explores the possible reasons exceptionally well. Two things that caught my eye:

- What is falling is the number of women who have > 0 kids. Once they have some kids, the number of kids per women is stabilized.

- And yes, the match between smartphones with high speed internet and declining birth rates.

But there's much more. Definitely recommended read to better understand on of the key problems of our time.

https://t.co/ysU8e4DK5o

[Jan Červinka @JanCervinka]

We started AI adoption with a central team collecting automation requests, teaching and helping build stuff. But it’s time to change that.

Our central team's job now is to remove what's getting in the way — tooling, data access, permissions, time. Whatever's blocking people from building, we deal with it. "Tell us what the roadblocks are" is the actual ask we put to people. Not "we are here to build it for you."

Some teams pick it up fast and ship. Others want to be told what to do. But activities like our hackathons are great for this. It seeds teams with AI-pilled members that start building and showing others what’s possible in their domain. It’s exciting to see the energy this brings.

[Jan Červinka @JanCervinka]

Skills, custom prompts, training — switching costs compound fast. Even with 5.5 looking exciting I'm still on current Opus, the friction isn't worth it.

No surprise both anthropic and openai invest heavily in enterprise adoption.

https://t.co/YmEp9shvfm

[Jan Červinka @JanCervinka]

Claude is excellent at building websites, apps, and slides. But when I tried to make a beautiful diagram, it stumbled. I expected more.

Tried Pencil, web-based editors, Mermaid. Nothing worked well. Then D2 saved the day. Not amazing, but good enough.

This discourages people. You hit something simple that turns out to be unexpectedly tricky with AI, and you want to drop it. Pushing through is where the learning actually happens.

That was Friday, a couple of hours before Claude Design launched. I was curious to see how much better it would do. After playing with it today, not much.

As always, I fall back to just Claude Code in the terminal. That said, I see real potential in Claude Design. The structured Design System prep is a good primitive for AI collaboration.

Anyway, what's your stack for beautiful diagrams with AI?

[Jan Červinka @JanCervinka]

@kevinrose Yes! I used to ask - “remind me, where were we?” Not anymore.

[Jan Červinka @JanCervinka]

This. I think catering to builders and entrepreneurs is one of the biggest opportunities right now. Everyone's a builder now.

[tobi lutke @tobi]

The most future-proof job: Entrepreneurship https://t.co/NbMUHPSoEi via @Shopify

[Jan Červinka @JanCervinka]

As I moved out of daily development into more managerial roles, I thought I'd be writing emails most of the time. Not code.

Well, AI changed that quite a bit. I'm still not writing code — but Claude is. Multiple things at once. While I'm on Google Meet.

Turns out my MacBook Air with 16 gigs of RAM wasn't built for that. Got the M5 Pro yesterday, and moving everything over became a pain.

Then I realized I could enlist Claude to help. Switched on SSH on the old Mac, asked Claude to connect, map the landscape, and carry over what matters to me now. Got Claude Code on the new Mac talking to Claude Code on the old one.

Surreal and fun. Next time, this is where I'll start — though I expect this will all be quite different when next time actually comes.

[Jan Červinka @JanCervinka]

Thats actually exactly why I prefer just a perfected Claude Code setup over OpenClaw. I feel I control the proces closer and focus on it. Telegraming my agent never felt good enough. Though the magic of having an agent with personality is lost with claude code

[Big Brain AI @realBigBrainAI]

Peter Steinberger, creator of OpenClaw, on why AI agents still produce "slop" without human taste in the loop:

"You can create code and run all night and then you have like the ultimate slop because what those agents don't really do yet is have taste."

Peter is direct: raw capability without direction still produces mediocre output.

"They are spiky smart and they're really good at things, but if you don't navigate them well, if you don't have a vision of what you're going to build, it's still going to be slop. If you don't ask the right questions, it's still going to be slop."

Great AI-assisted work is defined by the human guiding it.

@steipete describes his own creative process when starting a new project:

"When I start a project, I have like this very rough idea what it could be. And as I play with it and feel it, my vision gets more clear. I try out things, some things don't work, and I evolve my idea into what it will become."

Most people skip this part entirely, front-loading everything into a single prompt and wondering why the result feels hollow.

"My next prompt depends on what I see and feel and think about the current state of the project."

Each step informs the next. The work itself is the feedback loop.

"But if you try to put everything into a spec up front, you miss this kind of human-machine loop. And then I don't know how something good can come out without having feelings in the loop — almost like taste."

The agentic trap is what happens when you remove yourself from the process too early.

[Jan Červinka @JanCervinka]

Breaking a reasonably protected system used to require someone skilled and motivated enough to dig. Most systems were safe by friction alone.

Claude Mythos (however capable it turns out to be) is a reminder that friction is disappearing fast. The 27-year-old OpenBSD bug needed 1,000 parallel agent runs — but it still found something that survived decades of human review. Next generation, it probably won't need the 1,000 runs.

We started running the best publicly available models against our own systems continuously. Basically an automated pentest. Because the alternative is letting less skilled, less motivated people find the issues first.

[Jan Červinka @JanCervinka]

Several of our senior people are shifting from managing teams to managing AI agents. Becoming key contributors again instead of just coordinators.

That transition needs protected time. But projects without a clear owner keep creeping in. They sound simple — "just a quick setup," "someone will handle it" — so nobody properly owns them. They accumulate and steal focus from the work that actually matters.

I decided to sit down with some of our key people to realign once AI initiatives landed on everyone's plate. Cut meetings. Split responsibilities. Made explicit who owns what and how much capacity they actually have.

Everything is possible now — AI made it all feel two prompts away. But doing things well still takes time and focus. Picking what not to do is the hard part.

[Jan Červinka @JanCervinka]

I can realistically run about 3 things in parallel with AI coding agents. More than that, and some get stuck because I can't guide them properly.

The bottleneck isn't the agents. It's my ability to maintain mental models of what each one is doing and intervene at the right moments. Three workstreams and I'm sharp. Four or five and something sits idle for 20 minutes because I forgot it was waiting on a decision.

Most AI content focuses on what the tools can do. What I keep hitting is how many things I can actually hold in my head at once.

Not sure if 3 is everyone's number or just mine.

[Jan Červinka @JanCervinka]

The scary thing is that AI tools turned everyone into half a developer. Non-devs installing random packages through Claude. The attack surface expanded to every employee with an AI coding tool.

[Andrej Karpathy @karpathy]

New supply chain attack this time for npm axios, the most popular HTTP client library with 300M weekly downloads.

Scanning my system I found a use imported from googleworkspace/cli from a few days ago when I was experimenting with gmail/gcal cli. The installed version (luckily) resolved to an unaffected 1.13.5, but the project dependency is not pinned, meaning that if I did this earlier today the code would have resolved to latest and I'd be pwned.

It's possible to personally defend against these to some extent with local settings e.g. release-age constraints, or containers or etc, but I think ultimately the defaults of package management projects (pip, npm etc) have to change so that a single infection (usually luckily fairly temporary in nature due to security scanning) does not spread through users at random and at scale via unpinned dependencies.

More comprehensive article:
https://t.co/EJAZbqAPIQ

[Jan Červinka @JanCervinka]

David Kasper keeps praising Pencil. Figma just released their MCP for design manipulation. So I compared them.

We had existing website and slide deck designs in Figma. I asked Opus to create a new page and a few slides from a markdown outline — through Figma MCP, Pencil MCP, and Pencil's built-in agents.

Both MCPs could do it but deviated from our brand significantly. Pencil agents were on another level. Faster, followed conventions exactly, mostly good to go.

If you're experimenting with AI-assisted design, give Pencil a try. Great job @tomkrcha.

[Jan Červinka @JanCervinka]

I'm thinking a lot about the balance between exploration and execution with AI tools.

It's easy to keep searching for the best workflows. The best tools. The best stack. It's fun. And it matters — you need to know what's out there.

But it's also where you can lose months.

Should we use Compound Engineering or Superpowers? Codex or Claude Code? Build an internal marketplace? Every week there's a new question. I catch myself doing this — exploring when I should be executing.

And it's not just me. It applies to the whole team. At some point you need enough people on the same setup that they can actually help each other. Share tricks. Build on each other's work.

Maybe something 10% better exists out there. But we lose more from everyone still messing around than from just picking one and committing.

It's time to pick one and get good at it.

How do you find the right balance between exploration and execution in times like these?