Start With a Prompt: Inside How Warp’s CEO Follows His Own AI Coding Mandate | First Round

Start With a Prompt:

Inside How Warp’s CEO Follows His Own AI Coding Mandate

10.8.2025

When Warp’s Zach Lloyd noticed senior engineers were reluctant to adopt AI tools, he rolled out a slightly more pointed new policy: Every coding task needs to start with a prompt.

AI mandates from CEOs are nothing new. But Warp’s founder goes further than most by actually following it himself.

“The most enthusiastic initial adopters of AI at Warp were designers, PMs and me,” Lloyd says. “I was a principal engineer at Google, and I still write code at Warp, getting about one coding change in a week, and I’m using this new generation of tools constantly. I personally only code by prompt now, but it’s a recent development.”

Lloyd says it’s only within the last couple months where the product itself has gotten good enough to build Warp within Warp. “We are now building much of Warp starting with a prompt these days, which is a bit wild since it’s over a million lines of Rust with a custom UI framework,” he says.

Warp’s business has transformed alongside Lloyd’s personal coding practices, evolving from reimagining the terminal to building what they call an “agentic development environment” (ADE). That places the devtool company squarely in the middle of the seismic shift that’s currently upending engineering workflows. “We went from building an experience for users to tell their computer what to do with commands in machine language, to letting them tell their machine what to do in natural language,” he says.

The impact has been explosive. “It wasn’t until we started really embracing agentic development that our revenue took off,” Lloyd says. “The first $1M in ARR took 300+ days, but we’re now adding $1M every 5-6 days, with revenue up 19x this year.” Warp users launch almost 3M agents per day, generating 250M lines of code weekly with a 97% acceptance rate of agent-suggested diffs.

But this isn’t a product or a revenue story. It’s a story about how Warp’s developers transformed their own internal practices — and about how Lloyd visibly drove that change from the CEO’s seat.

To further proselytize adoption, he builds in public for all to see — at All Hands, in customer calls, in prompt-based development tutorials, even on weekly YouTube livestreams. And as it turns out, in media interviews, too.

When we interviewed him recently on how they were using AI internally, Lloyd took the liberty of sending over a video he made just for Applied Intelligence readers, showing — not telling — how the “ How Warp Uses Warp to Build Warp” guidelines steer product development at the company.

Following his own coding mandate step-by-step, he built a fix for Warp’s tab renaming interface, speaking his prompts rather than typing them.

The (senior developer) resistance problem:

The mandate’s origin story stems from Lloyd’s observation that senior developers were the ones resisting AI most strongly. When Lloyd posted about this problem a few months ago, hundreds weighed in in the comments, with many confirming they were facing the same struggles in their own organizations.

“We are literally building an AI devtool at Warp and I still find it hard to get folks to change their habits by using these tools more.”

Zach Lloyd

Founder & CEO at Warp

It makes sense that non-engineers gravitate towards these tools most, because they make the impossible possible, he says. “One of our designers recently helped build the new version of the terminal multimodal input, and he didn’t handwrite a single line of code. The idea that our designers would be building features and fixing bugs on their own would have been very hard to imagine a few years ago.”

Whereas for senior devs, AI is at best a speedup. “On the surface, these engineers ‘need’ AI the least because they know what they are doing,” he says. They also have wildly different dispositions toward AI. “Developers are a very diverse group of people when it comes to how they like to work. I find these tools very fun to use personally, but some folks — if you look at this summer’s Stack Overflow survey — are pretty against AI on principle,” he says.

But the resistance can run deeper than just lingering skepticism. “When you first try to do things by prompt, there's a high likelihood it will fail, or actually take you more time,” Lloyd says. “It’s not like if you try it just once you’re going to be instantly converted. That’s why many senior devs tend to think the code these tools produce is not very good, and that they can do it faster themselves.”

The coding mandate:

Motivated to move the needle on internal adoption, Lloyd circulated the following guidelines, which are published in full here (along with several other Notion guides on how the Warp team operates internally). “If you’re a founder or eng leader trying to get your team to use agentic tools, you need to set very clear expectations, probably much more explicitly than you think. I had to push pretty hard. It’s the difference between success and failure,” he says.

We go over this mandate every week at our team meeting because we really want to get people in the habit of working in this new way, and emphasize over and over that it’s not going to work well initially, that there’s a whole bunch of skill in how you use it. The guidelines in our mandate will change over time as the models improve, but they represent our best practices today.”

“If you’re not already, use agents to start — but not necessarily complete — every single task. It will change your coding practice.”

Zach Lloyd, Founder & CEO at Warp

Here’s the quick snapshot:

Step 1: Every coding task should start with a prompt in Warp

“You better give people a reason to change their behavior,” says Lloyd. At Warp, he outlined three core reasons:

  1. Productivity: “It’s still a bit squishy to measure, but I’m convinced it’s actually faster for many tasks and should increase overall engineering output. The main reason here is it allows for multi-threaded development. “Especially for someone like me where coding isn’t my main thing anymore, I can now have multiple things going at once. I can literally be in a meeting and have the change I want to make running in the background,” says Lloyd.

  2. New skills: It’s especially useful when you’re trying a new skill or working on areas of code you are less familiar with. “Given how quickly our codebase grows, that’s likely a lot,” says Lloyd. Senior engineers have really come to like starting by prompt when they’re trying something new.

  3. Dogfooding: A not insignificant part of my reasoning for the mandate is to encourage everyone to use and give feedback on what we are building. **We are asking our users to work this way and betting our business on it, so we had better believe enough in the workflow that we are willing to use it ourselves,” he says.

After three months, it's starting to have an impact. “Every single programming test that we’re doing right now to build Warp is starting with a prompt in Warp. Not every task is finished by prompting, to be clear, but the completion rate is getting higher week over week. I’d estimate we’re around 40-50%,” Lloyd says.

Step 2: If the coding task succeeds, great! Share the success in #warped-it

Spinning up a Slack channel dedicated to AI wins has become standard fare across most companies by now. Explicitly making this behavior part of the coding mandate was important for Lloyd.

Coding by prompt is a totally new discipline and we all have different instincts on how to do it well,” he says. “I’m in the channel posting my own PRs when I can, but I’m always lurking to get a more informal feel for the progress we’re making in using these agentic tools.”

Step 3: If, after 10 minutes, you feel like you are wasting time by prompting…

  1. Please share feedback with the relevant #feedback-channel on what failed
  2. Please also share your prompt, conversation id, and so on so we can learn where the issue was and possibly build an eval around it
  3. This is also a chance for folks to give each other tips on how best to use Warp to code

Step 4: Try another AI coding tool (e.g. Cursor, Claude) and see if you have more luck — if you do, please report what went well and what didn’t.

Step 5: Finally, just code by hand…we need to actually ship software and if that’s still fastest, you should do it.

The nuance to making it work:

As with most things, the real meat is in the more nuanced guidelines further down Warp’s mandate doc. “These are all focused on making the code as high quality as possible,” he says.

Principle #1: Avoid outcome-based prompting

“This is a newer one we added in more recently,” says Lloyd. “After the mandate had been in place for a little while, I came to realize that we needed to more explicitly forbid what I call outcome-based prompting, where you basically describe how you want the app to work rather than how you want a feature to be built or a bug to be fixed. Devs need to say how a change should be made.”

Principle #2: You own the code. Period.

“I kind of hate the term ‘vibe coding,’” Lloyd admits. “It implies that when you code with AI you shouldn’t even bother understanding the code, and that the right way to develop is by just re-prompting until the thing works. This is fine for low-stakes projects, but definitely not for pro-development.”

Principle #3: Babysitting beats one-shotting

“When I first started making changes with prompts, I would do a thing where I would just prompt, not really look very closely at the code. You kind of have to babysit it on harder changes right now, and if you don’t do that, you run a high risk of putting a bunch of time in trying to get the thing to work, only to look at the code and be left feeling that it technically works, but it’s not really useful, or it can't be merged.”

The future of engineering:

While Lloyd’s pleased with the progress to date, he’s got more changes on deck for Warp’s eng team. In addition to tracking more concrete productivity measurements, he’s looking at implementing a new technical interview session. “We’ll likely ask engineering candidates to build something by prompt before the end of the year, we just haven’t spec’d it out or calibrated how to do that interview yet,” he says.

“In the past, we’ve asked them to come in with a couple of ideas on how to improve Warp, and what I want to see these days is for them to suggest problems that we could solve with agents rather than terminal UI or collaboration ideas.”

“The core of the job is shipping great software that solves user problems. For better or worse, coding is an implementation detail.”

Zach Lloyd