How Anthropic’s Claude Code Team Actually Works: The Fireside Chat That Changes What You Think About AI Agents
Anthropic's Claude Code team revealed at the AI Engineer World's Fair that Claude Tag now handles 65% of their internal product engineering pull requests and that heavily scripted prompts are no longer best practice for the latest models. Simon Willison's detailed transcript of the session offers non-technical professionals a clear window into how multiplayer AI collaboration is reshaping knowledge work.
At the AI Engineer World’s Fair earlier this month, Simon Willison hosted a fireside chat with Cat Wu and Thariq Shihipar from Anthropic’s Claude Code team. The full video and an edited transcript are now available on Simon Willison’s site at simonwillison.net — and if you work in any knowledge profession in India, the conversation contains several signals worth paying close attention to, even if you have never written a single line of code.
This post unpacks the most significant revelations from that session in plain language, with honest context about what these changes mean for you right now versus what is still evolving.
The Number That Stopped Everyone: 65% of PRs
The single most striking data point from the chat is this: Claude Tag, Anthropic’s new collaborative Slack integration, now handles 65% of the product engineering pull requests (PRs) for the Claude Code team itself. To translate that for a non-technical reader: a pull request is essentially a package of new work — a bug fix, a new feature, or an update — that a software team reviews and approves before it goes live. More than half of those work packages inside Anthropic’s own product engineering team are now being created and landed by an AI agent, not a human developer sitting at a keyboard.
Cat Wu describes this as “a huge shift” and frames it not as replacement but as reallocation — engineers are moving toward higher-level thinking about what to build, while the AI handles more of the implementation.
What Is Claude Tag, and Why Should Non-Technical Professionals Care?
According to the transcript, Claude Tag is Claude living inside your team’s collaboration tool — currently launched within Slack. Three properties make it different from simply pasting questions into a chatbot.
A Concrete Scenario: A Mumbai Marketing Agency
Imagine you run a twelve-person content and performance marketing agency in Mumbai. Your team uses Slack. A client in the fintech space wants weekly performance summaries, campaign copy variants, and competitive landscape notes — all in a consistent brand voice.
Today, that means your account manager chases the analyst, who exports data, who hands it to the copywriter, who revises after the brand review. Each handoff loses context.
With Claude Tag as described in the Willison transcript, you could add it to your client channel, instruct it once to pull from your event data store each Monday, generate a draft summary in your established tone, and flag it to the relevant team member. Your brand preferences, client quirks, and formatting rules are stored as team memory. A colleague can jump into the same thread to redirect Claude’s draft without starting a new conversation. The non-technical account manager, the analyst, and the designer are all in the same collaborative session — no one needs to know how to prompt Claude individually, because the context is shared and visible to everyone.
Thariq Shihipar specifically mentioned that Anthropic’s marketing team uses Claude Tag to ask questions about features — Claude clones the codebase internally and returns a description plus a recording of the feature in use. You do not need to be a programmer to benefit from this kind of capability; you just need to be in the channel.
The Prompt Engineering Reversal You Need to Know About
If you have spent time learning how to craft detailed prompts — loading them with examples, long lists of “don’t do X, don’t do Y” instructions — the Claude Code team’s internal experience should make you pause.
According to the transcript, adding examples to a system prompt is no longer best practice for models like Fable 5 or Opus 4.8. The Claude Code system prompt itself was recently reduced in size by 80%. Cat Wu and Thariq Shihipar note that long lists of negative instructions (“don’t do this, don’t do that”) can actually reduce the quality of results from the latest generation of models.
This is a significant reversal of advice that circulated widely in 2023 and 2024. The implication for you: if you are still using heavily scaffolded prompts you built twelve months ago, it may be worth simplifying them and testing whether the newer models perform better with less hand-holding.
The internal Anthropic term for dogfooding their own products, by the way, is “ant fooding” — a small detail from the transcript that reveals how seriously they take internal testing as a quality gate.
Rewrites Are Now Good. What That Means in Plain Language.
Thariq makes a point in the transcript that overturns one of the oldest rules in software: never rewrite from scratch. His position, informed by actually doing it, is that rewrites are now good — Anthropic has internally rewritten a runtime called Bun in Rust, and it is already live for them.
For a non-technical reader, the implication is not about code. It is about iteration speed. The underlying principle is that the cost of rebuilding something from scratch has dropped dramatically, which changes how you should think about any knowledge product you produce — a process document, a client proposal framework, a training curriculum. The willingness to tear it down and rebuild it cleanly, rather than patching it endlessly, is increasingly the right instinct.
Auto Mode and the Deep Blue Problem
Anthropic describes their “auto mode” — where the agent makes decisions without asking for human permission at every step — as an enabling technology for Claude Tag. The transcript notes that Thariq advises offsetting what he calls the “Deep Blue” feeling (the unsettling sense that the AI is doing your work) by “being more ambitious” with the work you take on. The answer to feeling displaced is not to do less; it is to aim higher.
Cat Wu’s framing is similar: the team has moved from closely monitoring every permission prompt to delegating implementation and focusing on “what is the right experience we should be providing to our users.”
Honest Limitations and What Is Still Unclear
Several important caveats are worth naming directly.
- Claude Tag’s 65% PR figure applies only to Anthropic’s internal product engineering team, not to all of Anthropic. The transcript makes this explicit. External figures for other organisations are not available.
- Claude Tag is newly launched. The fireside chat took place when it had been publicly available for roughly a week. The social dynamics of multiplayer AI collaboration — who steers, how conflicts are resolved — are, by Cat Wu’s own admission, still being worked out.
- Fable, the video-capable model Thariq references, is a product Thariq used to edit its own launch video. Details about its general availability, pricing (in INR or otherwise), or region-specific rollout are not covered in the transcript and should not be assumed.
- Claude Tag is currently launched within Slack. Whether it integrates with other tools used widely in India — Microsoft Teams, Google Chat, or regional platforms — is not addressed in the source.
- The prompt engineering guidance about reduced system prompts applies specifically to Fable 5 and Opus 4.8 class models. If you are on an older model tier, the older guidance may still apply.
What to Watch For
The fireside chat, as documented by Simon Willison, points toward a few clear trajectories worth tracking. Multiplayer AI collaboration — where a whole team steers the same agent session — is moving from experiment to default workflow inside one of the world’s leading AI labs. The quality gate for shipping features is shifting from specification documents to internal retention data: Anthropic only ships features that demonstrate genuine user retention among its own employees first.
If you are a non-technical professional in India — whether in marketing, finance, operations, or content — the practical starting point is not to build anything. It is to read the Willison transcript carefully, understand what Claude Tag’s proactive and multiplayer properties actually mean for team workflows, and identify one repetitive, channel-based task in your Slack workspace where a persistent, memory-enabled agent could reduce coordination overhead. The infrastructure is arriving; the more valuable skill right now is knowing which problem to point it at.
