Claude Mods In-Depth: Harness Shifts from Managing Machines to Managing Human Attention

Claude Mods in Depth: When the Harness Shifts from "Managing the Machine" to "Managing Human Attention"

From an article by @daniel_mac8 published on @latentspacepod, he argues: Claude Mods lets users and Claude together use code to reshape the human-facing layer of Claude Code, deciding when to interrupt you, what to show you, and which decisions to let you make. Dan sees this as an early form of an "attention interface," and the Harness may thereby evolve into a platform on which others can build products. https://t.co/O98W0TPMUg

What Mods Are · Essentially Claude Code plugins: a set of TS/JS functions that listen to events such as prompts, tool calls, and interface rendering, and can observe, rewrite rendering, or replace default behavior. The author calls them "middleware for the Harness," usable in both the CLI and the desktop app. · Claude Code has a built-in skill for writing Mods: describe your needs in natural language and Claude can write it for you. · Six Mods are already built in officially: agents-md, diff, plugin-authoring, sec-default, telemetry, you-should-know. In other words, part of Claude Code's own functionality has already been reimplemented as Mods.

Two Mods Dan Recommends · You Should Know (officially built in, off by default): a side agent watches the main agent work and flags information buried in the execution log that has a big impact on the outcome. Example: when the main agent was editing files, it warned the author that about 45 files had uncommitted changes, and that discarding this demo's edits afterward would also throw away the earlier work. · Next Steps Supervisor (one of the ModSmith plugins Dan open-sourced): triggered when a round has edited files, run commands, or called tools 8 or more times; it lists points that need attention and turns suggested next steps into buttons that send to Claude with one click.

Comparison with Two Others: The Core Difference Is Who Designs the Attention Interface · ChatGPT Intelligent UI: for each answer, the model picks and generates an interface from OpenAI's component library. Zero configuration, but very low controllability. · Cursor PR Routing & Approval: teams write policies, and a fixed process executes them, including automatically approving low-risk PRs by risk score. One design worth noting: if a PR modifies the approval policy, the system reviews that PR under the old policy on the base branch, preventing an agent from changing the rules to let itself through. · Claude Mods: users and the agent build the interface itself together in code, and it can be hooked into any node of the loop. The highest degree of freedom, at the cost of being responsible yourself for whether the interface is any good.

Problems Dan Himself Points Out · Noise: after installing it on October 15, the Supervisor triggered in 99 of 111 rounds. With reminders that frequent, people start skipping them. · Cost: each check forks the session and re-reads the context. Per the README's estimate, a 50k-token session consumes about 5–6k equivalent input tokens per check, and more once the cache expires. · Trust: Mods run with user permissions, with no sandbox; they can read and write files, see all prompts, and consume quota. Built-in Mods still run under --safe-mode.

Criteria and Direction Dan judges whether a Mod is useful by three things: whether it surfaces something you would otherwise have missed, whether it makes decisions easier, and whether it lets you act directly without being distracted by more output. The next idea is to have the Harness read session logs itself, spot places where you repeatedly correct the agent or repeatedly check manually, and then proactively suggest writing a Mod.