Claude Code and Claude Tag: Anthropic Redefines Engineering

A Fireside Chat with Cat and Thariq from the Claude Code Team

This month, during the AI Engineer World’s Fair, I had the opportunity to moderate an exciting discussion with Cat Wu and Thariq Shihipar, prominent members of the Claude Code team at Anthropic. We covered various topics, including Claude Code, Claude Tag, Fable, the safety of coding agents, performance evaluations, as well as the design and internal use of these tools by Anthropic.

The full session is now available on YouTube. What follows is an edited version of the transcript, enriched with additional links and my own comments in bold.

Key Takeaways

• Claude Tag, the new collaborative integration of Slack from Claude, now accounts for 65% of the product engineering PRs for the Claude Code team.

• Claude Code first rolls out features to Anthropic employees and only launches those that show user retention within this group.

• Critical changes to Claude Code are always manually reviewed, but the team is increasingly relying on automated code review for the product's "outer layers."

• Adding examples to a system prompt is no longer best practice for models like Fable 5 or even Opus 4.8. The system prompt for Claude Code has recently been reduced by 80%.

• Similarly, lists of "do not do X and do not do Y" can diminish the quality of results from the latest models.

• The term used at Anthropic to refer to self-use is "ant fooding."

• Anthropic firmly believes in their automatic mode, which they consider an enabling technology for Claude Tag.

• Thariq advises compensating for the undesirable effects of coding agents by being "more ambitious" in the work you undertake.

• Fable is capable of editing videos, and Thariq used it to edit his own launch video.

• Anthropic's internal work culture in public is essential to their success, as evidenced by their use of Claude Tag in their public Slack channels.

How Has Your Daily Work Changed Over the Past Year?

Simon: Claude Code was released in February of last year — it’s less than a year and a half old, and it was originally just a point on the launch of Claude Sonnet 3.7. How has your daily work changed over the past year, now that we have these coding agents actually working for us?

Cat: I remember when we first launched Claude Code and Sonnet 3.7, you had to give it a task and closely monitor every little thing it tried to do. I would read every permission prompt very carefully. I often said no — no, no, no, did you check this file? Did you check that other file? And now, it’s incredible with each model generation. I feel like we’ve all had the chance to step back and delegate much more of the mundane implementation to Claude. It has freed up a lot of time for us to think about more creative work, like: what is the right experience we should offer our users, now that we know Claude Code can implement a lot of it? And now with Fable, it’s a totally different step change. We see for many of our use cases that you can actually achieve a great number of features with Fable now.

Thariq: I remember the first message I got about Claude Code. One of my best friends told me, "You have to try Claude Code." It was around the time Opus 4 was released, and I tried it and thought, "Oh man. I need to work at Anthropic now." And it was Opus 4 — a great model, but you were reading permission prompts. It’s a bit crazy how we have amnesia, where I think, oh, the automatic mode has always been there, right? I don’t even remember hitting yes and allowing. For me, the big thing I’m trying to focus on is that we need to produce higher quality work than ever before. The results are incredibly high quality. I’ve used it to edit videos, and I think, okay, this has to meet our brand team’s very strict requirements in a few hours, or we just can’t do it. That’s how I’m trying to adapt with Fable: the best work we’ve ever done, faster than ever.

What Part of Conventional Software Engineering No Longer Holds?

Simon: What is one part of conventional software engineering that was true a year ago and you no longer think is valid in this new world?

Cat: One of the biggest changes we’re seeing in engineering skills: two years ago, it was quite typical for a product manager to go talk to a bunch of customers, align for six months with cross-functional teams on a PRD, and write a detailed specification on exactly how we’re going to implement this before the first line of code is written. Now, things are completely reversed. For many engineers, the encouragement I would give people in the room is to develop your business sense and product sense more on what we should build, because the time between having an idea and building it is so much shorter — it’s gone from six to twelve months to maybe even a week. This means we all need to have a better taste for what’s worth building, which will actually influence the businesses we’re working on. So, there’s an increase in the value of product taste and business sense, and a bit less on execution in most product areas. Of course, for infrastructure, there’s still a very strong emphasis on making sure all the details are correct.

Thariq: For me, it’s that rewrites are now good.

Simon: The worst thing you could do is now actually acceptable!

Thariq: Exactly. All the talk about the "Mythical Man-Month" — never rewrite — I’m now for rewriting. If you have a good test suite — and I think rewriting forces you to ensure you have a good test suite — but I think what people underestimate is that a codebase is a specification, and maybe it’s the only copy of the specification you have, because no one knows every part of the codebase. You can take that as an artifact and distill it or create other versions. We rewrote Bun in Rust and it works great — it’s online for me right now.

Simon: You’re not deploying Claude Code on Bun-in-Rust yet, are you?

Thariq: Internally, we have.

(In fact, it seems that Anthropic started deploying Claude Code on Bun-in-Rust for everyone on June 17.)

What Are Non-Engineers Doing with Claude Tag?

Simon: The other big launch recently was Claude Tag — it’s been what, a week now, at least for the rest of us. I understand it’s being widely used at Anthropic by non-engineers. What are non-engineers doing with Claude Tag?

Cat: Claude Tag is a Claude that lives in your team’s collaboration tools. We launched it last week in Slack. The difference with Claude Tag is that it’s multiplayer by default. Once you add Claude Tag to a Slack channel, you can chime in, your teammates can chime in, and you can collaborate together on the PR. The other big difference is that it’s proactive rather than reactive. You can tell Claude Tag: "Hey, monitor every bug report in this channel, set up a PR to fix it, and tag the engineer who touched that part of the codebase," and it will do that for the entire lifetime of the channel without you having to tag it manually. And the third big change is that we’ve added team memory to this. If you tell Claude Tag your preferences in the channel, it will remember them for every future message. If you always want it to debug crashes but don’t want it to debug warnings, just say that in natural language in the channel and it will remember it for you and all the other members of your team.

Internally, we see Claude Tag as the evolution of Claude Code. We see this as a big change in how we work internally. Claude Tag currently represents 65% of our product engineering PRs.

Simon: For all of Anthropic, or just for Claude Code?

Cat: This is only regarding our product engineering team — our internal version of Claude Tag represents 65% of our product PRs right now. And that’s a huge change; it’s more than 50% of our PRs. The way we see people dividing the work between Claude Code and Claude Tag is as follows: Claude Code is still the best place for your most complex tasks, when you’re iterating interactively with the agent. But Claude Tag is great for working proactively on your behalf, so you no longer need to manually launch Claude Code for all the bug reports that come up for the features you’re working on.

Thariq: And for non-coding cases: for example, before this discussion, we asked Claude Tag: "Hey, when will Fable be launched?" We wanted to make sure to coordinate it with the announcement. Claude Tag would search our Slack and see who said what. As a search engine for your company, it’s really valuable. It has all the context for your product, so you can ask it questions related to metrics — often when you’re making decisions, you want them to be informed by what the metrics say, so you connect it to your event store. I’ve seen our marketing team do things like: "Hey, tell me about this feature." They’re not programmers, but Claude is a programmer — it can clone the codebase and say: "Here’s the feature, here’s what it looks like, here’s a recording of me using the feature." This allows for a wide variety of things, and I think we’re still at the beginning of discovering this.

Claude Tag as a Collaborative Team Layer

Simon: One of the issues I’ve encountered with coding agents is that I understand how to use them as an individual, but I’m not really clear on how to use them in a team environment. It seems that Claude Tag is your current answer to this collaborative team layer for these tools.

Cat: Exactly. And a large percentage of our sessions are actually multiplayer right now. Maybe I say: "Hey, I think we should implement this new feature in Cowork," and I’ll tag Claude Tag to do a first pass. Then I’ll tell Claude Tag: "Share a recording of your final implementation," and I’ll tag design to take a look. They’ll adjust it, then pass it to engineering to finalize and deploy it.