Claude Tag closes 65% of Anthropic's PRs as system prompts slim 80%

Claude Tag, Anthropic's Slack-native coding agent, handles 65% of the product-engineering pull requests for the Claude Code team, prompting a reduction of 80% in the system prompt's mass. This indicates that frontier models require less scaffolding. The agent operates within Anthropic's public Slack channels, with an "auto mode" that eliminates the need for human approval of each tool call. In a fireside chat transcribed by Simon Willison, Claude Code engineers Cat Wu and Thariq Shihipar detailed a workflow where Claude Tag generates diffs openly, automates code review for the "outer layers" of the product, and reserves human reviewers for critical changes. The models have shifted toward Fable 5 and Opus 4.8, where the team found that example-heavy system prompts and negative constraints degraded quality. The Claude Code system prompt was reduced, abandoning few-shot prompting in favor of sparse, direct instructions.

At Anthropic, internal testing is referred to as "ant fooding," acting as a live retention gate. Features are shipped to employees first, and only those that gain traction internally are rolled out broadly. This has reduced the idea-to-ship timeline from six to twelve months to about one week, prioritizing product taste over raw implementation skill. Shihipar advocates for rewriting code, arguing that a disciplined test suite makes rewrites safer than preserving legacy code, especially when the codebase is the only specification document. Monitoring has shifted; Wu has moved from scrutinizing every permission prompt to delegating broader authority to the agent.

The team did not release inference economics, such as dollars per million tokens, latency for auto-mode tool chains, GPU-hour burn rate, or throughput ceiling. The 65% PR share figure is unqualified—it's unclear if it reflects net velocity gain or PR volume inflation, whether automated review maintains defect rates as submission volume rises, or the token cost saved by the slimmer prompt at scale. Anthropic also did not specify how many merged PRs required human amendment afterward.

The trust boundary is the most significant unsized risk. Anthropic has not published the criteria that distinguish routine changes eligible for automated review from "critical" changes requiring human review, leaving replication teams to guess where bottlenecks or regressions may occur. The prompt reduction introduces a migration challenge: any production system stabilized by few-shot examples and negative-constraint lists will face a breaking re-architecture to move to Opus 4.8-class models, as the scaffolding that supported older generations now hinders new ones. The 65% figure comes from the team building the agentic tool itself; the adoption curve for other teams is unreported.