Dan McAteer, an AI engineer who writes for Latent Space, spent a few weeks bending Claude Code to his own workflow with Claude Mods. His claim: “Claude Code is great out of the box, but every engineering workflow is different. Claude Mods let you make Claude Code yours, in the way you want to pay attention.”
What a mod is
A mod is a Claude Code plugin made of TypeScript or JavaScript functions that run inside the tool. McAteer describes them as responding to events such as a prompt, a tool call, or the screen being drawn. A mod can watch an event, change how it appears, or handle it instead of the default behavior. In his words, mods are “middleware for the agent harness.” The harness is the layer of software around a model that runs its tools and talks to you.
His broader thesis is that as models absorb more of the harness, what remains faces the human. He calls that the “Attention Interface”: the layer that decides when the agent interrupts you, what it shows you, and which decisions it needs from you. His bet: “ultimately, the harness becomes a platform that other companies build on, with models as the engine.”
A working example, and what it cost
McAteer built a “Next Steps Supervisor” mod. It runs after a turn that edited files, ran commands, or made eight or more tool calls, then suggests next steps as clickable buttons.
He reports three problems. It was noisy: 99 of 111 turns triggered it. It cost tokens, which he estimates at 5 to 6 thousand input-equivalent tokens per check in a 50,000-token session. And mods are not sandboxed, so they can read and write your files and see every prompt you send. His test for keeping a mod: “did it surface something I would have missed, make a decision easier, or let me act without getting distracted by more output?”
A mod that fires on nearly every turn is another thing to ignore
The strongest part of the post is that a mod’s cost is measured in attention, not only tokens. McAteer’s own number shows the Attention Interface only works if the trigger rule is selective.
The platform claim is a prediction. A few weeks of personal use by one mod author does not show that other companies will build products on the harness. The security point matters more today. Anthropic has said mods run with the same access to your machine as Claude Code, so installing someone else’s mod means trusting code that can see everything you type.
If you write a mod, trigger it on a narrow condition. Count how often it fires and what each check costs before you keep it.
End of article