Claude Code now has a built-in way for users to rewrite how it behaves on screen: mods, small pieces of TypeScript or JavaScript that run inside the app and change what it does and how it looks. Dan McAteer, writing on Latent Space, spent several weeks shaping Claude Code around his own engineering habits and argues that the real contest in coding agents is shifting from what the AI can do to what it asks of you.
A mod listens for events, such as a prompt being sent, a tool being called, or the screen being drawn. It can watch the event, change how it appears, or take over from the default behavior entirely. McAteer calls them middleware for the harness, the software wrapped around the model that lets it edit files and run commands. You can also just ask for a mod in everyday words, because Claude Code ships with a skill that lets it write them for you.
McAteer’s larger idea is what he calls the Attention Interface. He argues that models keep pulling harness duties into their own weights, so what remains is the layer that faces the human: when the agent interrupts, what it shows, which decisions it hands back. Mods are an early, user-editable version of that layer. The term and the argument are his, set out in an earlier Latent Space post, not an Anthropic slogan. The mod system itself is Anthropic’s, and the company has shipped six built-in mods, including one that loads project instructions and one that takes over the diff view.
His favorite is a built-in called You Should Know, which is off until you switch it on. A side agent watches the main one and puts a note over the input box if something important is buried in the trace. In his example, Claude Code was trimming interface text in a personal project when the note warned that the file already held uncommitted work, along with about 45 other files. Discarding the demo edits would have thrown that work away.
The second mod is one McAteer built himself, Next Steps Supervisor, part of an open-source plugin he calls ModSmith. It fires after any turn that edits files, runs commands, or uses at least 8 tools, and lists what was flagged and offers a clickable next step. He compares his approach with two other products. OpenAI’s Intelligent UI, launched with GPT-6 on October 7, lets ChatGPT generate charts, forms, and small tools from a component library OpenAI built. Cursor’s PR Routing & Approval sends pull requests to reviewers by code ownership and risk score, and reads approval policy from the base branch so an agent cannot loosen the rules for its own change. In McAteer’s framing, those two give you an attention design someone else made. Mods let you build your own.
He is frank about the costs, and they matter for anyone tempted to copy him. Since installing the supervisor on October 15, 99 of his 111 turns triggered it, so it speaks almost every time and he finds himself skimming. By the README’s own estimate, each check costs 5 to 6 thousand tokens of input on a 50,000-token session, more if the cache has expired. And Anthropic’s docs, he notes, say a loaded mod can open and change your files, see every prompt, and spend your usage, and that mods are not sandboxed. Built-in mods keep running even in safe mode.
His own test is blunt: did the mod surface something he would have missed, ease a decision, or let him act without extra distraction? Everything else is window-dressing. Because the numbers come from one engineer’s 111 turns, they show how a mod can backfire, not how often it does.
For teams standardizing on Claude Code, the practical question is governance: an unsandboxed plugin that sees every prompt deserves the same code review as any other dependency, and the first mod worth auditing is whichever one a colleague writes to be helpful.
Reported by Latent Space (Dan McAteer, “Claude Mods: Flipping the Harness Inside Out”); the source page carries no publication date.
