Claude Mods are plugins for Claude Code implemented as TypeScript or JavaScript functions that run inside Claude Code. They customize the behavior and appearance of the CLI and Desktop interfaces by responding to events — prompts, tool calls, or interface renders. A mod can observe an event, alter how it’s rendered, and handle it instead of the default behavior.
Why they matter: the Attention Interface
The author frames Claude Mods as an early form of the Attention Interface — the layer that decides when an agent interrupts you, what it shows, and which decisions require human judgment. Mods act as middleware for the agent harness, injecting code that changes how Claude Code executes and renders, and making the human-facing part of the harness programmable by the user.
Reversing the harness
Originally, the agent harness served as an interface between the agent and the machine it runs on: you prompt the agent, it runs commands, edits files and calls tools. The article argues that the harness is becoming the channel by which the agent interfaces with the human. Claude Mods let you flip the harness so it points from the agent to the human operator, turning the harness into a programmable way to manage human attention.
Built-in mods and practical examples
Anthropic ships several built-in mods with Claude Code; the author lists six available under /plugin > Installed > Built-in:
- agents-md: loads AGENTS.md as project instructions.
- diff: takes over /diff and draws its pane.
- plugin-authoring: gives Claude the skill it uses to write mods for you.
- sec-default: enforces organization-managed constraints against user-installed mods.
- telemetry: sends analytics records from Claude Code and built-in mods.
- you-should-know: a side agent that shows a note above the prompt when it finds something you might have missed (off by default).
The author reports enabling You Should Know and gives a concrete example: while simplifying UI text, the main agent edited CompareView.swift and installed a new build, but the You Should Know mod flagged that CompareView.swift and roughly 45 other files had uncommitted work. The mod’s suggestion was to commit or stash prior edits first; following that advice prevented accidental loss of earlier work.
The author also describes a mod they built as part of an open-source project called ModSmith: Next Steps Supervisor. This mod runs only after a turn that edited files, ran commands, or made eight or more tool calls. Since the author installed it on October 15, 99 of their 111 turns in Claude Code met that threshold. The mod highlights multiple items for attention and offers a suggested next step as a clickable button that sends an instruction to Claude Code.
Comparable approaches: OpenAI and Cursor
The article compares Claude Mods to two other attention-shaping systems:
-
ChatGPT Intelligent UI (OpenAI): On October 7, OpenAI shipped GPT-6 inside ChatGPT and introduced an "Intelligent UI" where the model can generate interface components (charts, maps, checklists, forms, small tools) on the fly as part of its answer. The model chooses which interface components to show from an OpenAI-provided component library. This provides zero-setup usability but little control for users.
-
Cursor PR Routing & Approval (Cursor): This system targets team workflows and streams of pull requests. It routes reviewers based on code ownership and commit history, can auto-approve low-risk PRs, and weighs a risk score against policy thresholds, Bugbot/Security findings and policy files in the repo. One explicit safety rule: if a PR edits an approval policy, the agent will not use that edited policy to relax review on the same PR; it reads the policy from the base branch.
All three systems shape what gets human attention and how; their key difference is who designs the Attention Interface. ChatGPT’s Intelligent UI picks interfaces answer-by-answer inside a lab-built frame; Cursor enforces team-written policies through a fixed workflow; Claude Mods let the user and agent build the interface in code, but the user is responsible for making it useful.
Downsides observed by the author
The author notes three practical downsides:
-
Noise. The Next Steps Supervisor condition (file edits, commands, or ≥8 tool calls) is met very often in practice; since installation on October 15, 99 of the author’s 111 turns triggered it. When a mod speaks up on almost every turn, users may start skimming past it, undermining the Attention Interface’s purpose.
-
Cost. Each supervisor check forks the conversation and re-reads it. The mod’s README estimates about 5–6k input-equivalent tokens per check for a 50k-token session, and more when the prompt cache has expired.
-
Trust. A mod runs with the user’s permissions. Anthropic’s documentation states that a loaded mod can read and write files, see every prompt you send, and use your usage quota; and "mods aren’t sandboxed." Built-in mods can also run when Claude Code is started with --safe-mode. The author is comfortable using mods they wrote themselves or that ship with Claude Code, but would inspect third-party code before installing it.
The counterpoint is that mods are configurable: if a supervisor is noisy for a session, /supervisor off disables it. Because users can change their mods, they can tune when and how a mod runs.
Where this could go next
Customizing the harness will not benefit every engineer equally: each added mod is another information source demanding attention. The author’s simple test of a mod’s usefulness is whether it surfaced something they would have missed, made a decision easier, or let them act without distraction; otherwise it’s just window-dressing.
Today, building a mod tends to work like this: you notice friction in your workflow, describe the mod you want, and Claude helps you build it. The author argues the next step is for the harness itself to notice recurring friction by analyzing session transcripts, identify repeated corrective steps, propose a mod, help test it, and refine it with user feedback. That would let the harness suggest improvements based on real-world use, which the author calls the best-case scenario for AI engineers. Claude Mods are positioned as a move in that direction.
Final thought
If the author’s bet is correct — that the harness becomes a platform other companies build on, with models as the engine — then in a few years we will discuss startups built on Claude Code, Codex, or yet-to-exist harness platforms. The article closes by asking readers what friction they’ve noticed while working with Claude Code that could be solved with a mod today.
