A Fireside Chat with Cat and Thariq from the Claude Code team
Source
Simon Willison
Author
Simon Willison
Date
Key takeaways · AI-distilled
Claude Tag is multiplayer and proactive by design. You add it to a Slack channel and give one standing instruction, like monitor every bug report, open a PR, and tag whoever last touched that code, and it keeps doing that for the life of the channel.
Preferences live with the team, not one person's config. Tell Claude Tag in the channel that it should debug outages but ignore warnings, and it applies that for everyone on the team from then on.
The team's split: Claude Code for the hardest tasks, where you want to iterate with the AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → turn by turn. Claude Tag for work you want done on your behalf without kicking off a session at all.
Thariq's case for rewrites: a codebase is a spec, and often the only complete copy of it, since nobody knows every branch. That makes it an artifact you can distill into a new version. Anthropic rewrote Bun in Rust and now ships on it.
The engineering agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition → mix shifts as idea-to-shipped collapses from six or twelve months down to about a week. Product taste and business judgment rise in value, execution falls, except in infra where the details still decide everything.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
Why it matters
Practical prompting guidance from the people who ship Claude Code — that adding examples and long 'don't do X' lists now hurt results on frontier models, and that their system promptThe standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.Full definition → shrank 80% — plus a candid look at how Anthropic gates features on internal retention and reviews agent-written code.
Key quotes
“Claude Tag currently lands 65% of our product eng PRs.”
Cat Wu
“For me, it's that rewrites are now good.”
Thariq Shihipar
“In general, we are trying to move to a world where humans don't need to be in the loop.”
Cat Wu
“The way you offset that is by being more ambitious.”
