# Anthropic's Claude Code Team Cut Its System Prompt by 80%
**Cat Wu and Thariq Shihipar say trimming examples and "don't do X" rules cut the Claude Code system prompt to one-fifth its old size.**

## Fewer rules, better answers

At a fireside chat hosted by Simon Willison at the AI Engineer World's Fair, Anthropic's Cat Wu and Thariq Shihipar described how prompting habits that helped older models are now backfiring on newer ones, as laid out in the [published transcript](https://simonwillison.net/2026/Jul/21/cat-and-thariq). For models like Fable 5 and Opus 4.8, adding worked examples to a system prompt is no longer considered best practice, and listing "don't do X and don't do Y" instructions can actually make results worse. Acting on that finding, the team shrank the Claude Code system prompt by 80%, cutting it down while keeping the remaining instructions direct and example-free.

## Claude Tag takes over the pull request queue

The pair also described Claude Tag, Anthropic's Slack-based coding collaborator that lives inside a team channel, watches for bug reports, and opens pull requests on its own. Internally, Claude Tag now lands 65% of the product engineering team's pull requests, more than half of everything that ships. Cat described the tool as "multiplayer by default," meaning any teammate can chime into the same thread, and it remembers stated preferences for the rest of a channel's life. Anthropic still relies on Claude Code itself for the most complex, interactively steered tasks, saving Claude Tag for lower-touch or proactive fixes.

## Rewrites beat refactors, but humans still check the risky parts

Thariq argued that full rewrites, once considered wasteful compared with incremental refactoring, are now often the faster and safer choice, as long as a solid test suite forces the rewrite to prove itself. Anthropic already runs an internal Bun-in-Rust rewrite of parts of its own tooling, and it reached everyone outside the company by June 17th. Even with that appetite for rewrites and automated review of the product's "outer layers," Cat and Thariq said critical changes still go through manual human review before shipping. Anthropic calls its internal dogfooding practice "ant fooding." Cat and Thariq said engineers used to check every permission prompt by hand but no longer have to, thanks to auto mode, which the team sees as the technology that enables Claude Tag.

## Key facts

- Claude Tag currently lands 65% of the Claude Code team's product engineering pull requests.
- The Claude Code system prompt was reduced in size by 80% after the team stopped adding examples and negative instructions.
- Fable 5 and Opus 4.8 were both named as models that perform worse with example-heavy or "don't do X" prompts.
- Anthropic's internal Bun-in-Rust rewrite of Claude Code reached everyone outside the company by June 17th.

## Check your understanding

What percentage of the Claude Code team's product engineering pull requests does Claude Tag currently land?
Why did Cat Wu and Thariq Shihipar say the Claude Code system prompt got 80% smaller?
Which two models did the team name as performing worse with example-heavy system prompts?

## Sources

- [A Fireside Chat with Cat and Thariq from the Claude Code team](https://simonwillison.net/2026/Jul/21/cat-and-thariq)

---

- Category: learn (skill)
- Affects: agent builders · prompt engineers
- Published: 2026-07-21T12:54:02.000Z
- Original: https://simonwillison.net/2026/Jul/21/cat-and-thariq
- Generated by tinker news