Break It Down

Break It Down is AR Labs' open-source slash command for Claude Code, Anthropic's command-line coding agent (an AI agent that writes code). It is for anyone who works with an AI assistant and finds its long answers hard to take in. When an AI assistant is about to answer with five questions at once, 10 options or a dense multi-part analysis, the command makes it break that answer into pieces and present them one at a time, waiting for the person's reply before moving on. It was released as open source under the MIT licence in December 2025, and it rests on a finding from cognitive science: people can hold only about four chunks of information in working memory at once.

What it does

Without the command, a planning or debugging session with an AI assistant often produces a wall of text: several questions, an analysis and a recommendation, all needing a response. The person has to hold all of it in mind to answer any of it.

With the command, the assistant first writes its pending response into a task list, one item per chunk. It then shows only the first chunk, waits, marks that chunk done when the person has answered, and moves to the next. The pattern is chunk, present, wait, repeat, until the list is finished.

flowchart LR
  D["Dense response pending"] --> L["Split into a task list"]
  L --> P["Present one chunk"]
  P --> W["Wait for the reply"]
  W --> C["Mark it done"]
  C -->|"more chunks"| P
  C -->|"list finished"| E["Done"]

It is meant for moments where the cost of overload is high: answering several questions about a design, weighing many options, working through a long code review section by section, or a planning session that needs a decision at each step.

Why it works

The repository explains the reasoning and cites its sources. Working memory is limited (Baddeley and Hitch, 1974), to roughly four chunks (Cowan, 2001), and exceeding that limit degrades decisions. People handle sequential, focused attention far better than parallel streams of information. Checkpoints with confirmation stop a misunderstanding early, before later steps build on it. The command applies these to the shape of an AI conversation: a dialogue with checkpoints instead of a monologue.

Richard M. Thompson, who founded AR Labs, explored the same interest in attention and working memory in his earlier book manuscript, Cognitive Ecology, and the "one question at a time" rule appears again in our interview skills and voice agents.

What exists today

The public repository on GitHub contains the command file, installation and usage guides, a background document on the cognitive science, and worked before-and-after examples. Installing it means copying one markdown file into Claude Code's commands folder. The pattern also works with other assistants that can keep a task list, and the README invites adapters for them.

This page covers the open-source release. How we use the command in our own sessions is described on the Break It Down skill page, and the idea behind it on the cognitive chunking concept page.