Claude Mods introduce programmable human-facing agent interfaces

Claude Mods are plugins for Claude Code that let users customize how the agent surfaces decisions, interruptions, and rendered outputs to the human operator. They invert the traditional agent harness so that it faces the human rather than the machine, turning the harness into what the author calls an "Attention Interface."

Anthropic has shipped six built-in mods, including "you-should-know," a side agent that flags material changes or decisions that might otherwise be buried in the agent trace. Another built-in, "diff," takes over the /diff command and renders a dedicated pane. Anthropic also bundles a mod that teaches Claude Code how to build new mods on request.

Users can build their own mods in TypeScript or JavaScript. These mods can observe events such as prompts, tool calls, or interface renders; modify how they are displayed; and handle them in place of default behavior. The author highlights two practical examples: the built-in "you-should-know" mod, which prevented unintended loss of uncommitted work, and a custom "next steps supervisor" mod that suggests follow-up actions after complex turns.

The article frames this as part of a broader trend: as models absorb more of the harness into their weights, the residual harness layer becomes the Attention Interface that decides when to interrupt, what to show, and which decisions require human input. The author speculates that the harness could evolve into a platform others build on, analogous to how the browser became a platform for web apps.