Anthropic ships Claude Mods to customize Claude Code

Claude Mods arrive as the agent harness is shifting from an interface between the agent and the machine toward one between the agent and the human. That shift traces to a simple observation: as models absorb more of the harness into their weights, what remains faces the person — deciding when the agent interrupts, what it shows, and which decisions it needs. Mods are an early, user-shapeable version of that layer, and some of Claude Code's own features already run as built-in mods, including agents-md, diff, plugin-authoring, sec-default and telemetry. The author's own experience shows both the promise and the friction: the You Should Know mod caught uncommitted work in CompareView.swift and about 45 other files before demo edits could discard it, while the self-built Next Steps Supervisor ran on 99 of 111 turns since October 15 and drew an estimated 5–6k input-equivalent tokens per check on a 50k-token session. The test for any mod is whether it surfaced something missed, made a decision easier, or let the user act without more output — otherwise it is window-dressing. The next step the author envisions is a harness that reads session transcripts to spot repeated friction and suggests a mod for it.