Skip to main content

04.16Conversation design

Break It Down: Claude Code command to chunk long AI replies

Open-source Claude Code slash command from AR Labs that turns an overwhelming AI reply into one question or chunk at a time, sized to human working memory.

Source on GitHub
Engraved plate: a ticket dispenser with a long coiled roll issues a single ticket marked one of five.An engraved ticket dispenser holds a long coiled roll; a crank on its side issues a single ticket marked one of five, waiting to be answered before the next.AR LABS · INSTRUMENT CABINET · DRAWER 04CALL NO. 04.1601234567891 OF 5one questionABCDFIG. 16 · BREAK IT DOWNTICKETDISPENSERPARTSAThe overwhelmingreply, coiledBWaits for youranswerCOne question orchunk at a timeDSized toworking memoryONE CHECKPOINT AT A TIME
On this page
  1. How it works
  2. The judgement it encodes
  3. What exists today
  4. Example

Break It Down is an open-source Claude Code slash command, /breakitdown, that AR Labs wrote to turn an overwhelming AI reply into a conversation. Use it when an agent is about to deliver five questions at once, a dense multi-section analysis or a long list of options: the command has it split the material into small checkpoints and present them one at a time, waiting for an answer before moving on. It applies a finding from cognitive science: a reply can be complete and still be more than a person can take in at once. The idea behind it is set out in cognitive chunking for AI output, and the release itself has a project page.

How it works

The command uses Claude Code's built-in task list (the to-do list the agent can create and tick off) to hold the conversation's plan, so the agent keeps track of where it is while the person sees only the current step.

FIG. 01
flowchart LR
  A["Agent about to overload"] --> P["Plan checkpoints in a task list"]
  P --> O["Present one checkpoint"]
  O --> W["Wait for the answer"]
  W --> M["Mark done, move on"]
  M --> O
  1. Stop and assess. Before replying, the agent asks itself whether it is about to ask several questions, present several analyses or options at once, or deliver a monologue where a dialogue would work better.
  2. Plan the checkpoints. It breaks the intended reply into sequential chunks: a confirmation, a clarification or choice, the next piece of information, a final recommendation.
  3. Execute one at a time. It marks the first chunk in progress, presents only that, waits, marks it complete, and continues.

The judgement it encodes

Respect working-memory limits. People hold about four chunks of information in mind at once, give or take one (Nelson Cowan, The magical number 4 in short-term memory (Behavioral and Brain Sciences, 2001)). Past that limit, decisions get harder and some questions go unanswered.

One question per turn. A good chunk asks one question, presents one analysis or gives one recommendation, and creates a natural stopping point.

Find the smallest coherent unit. Some information must be presented together. When chunks depend on each other, the agent bundles them and says why.

This is not dumbing down. The full content still arrives; it arrives in an order a person can respond to. If someone asks for everything at once, they get it, with an offer to break it into steps.

Catch it mid-flow. If the agent notices it is already halfway through a long dump, it says so, offers to switch to sequential questions, and restarts from the first checkpoint if the person agrees.

What exists today

The command is a single markdown file, public on GitHub under the MIT licence; installing it means copying that file into Claude Code's commands folder, as the README describes. We still use it, and the same principle runs through several other skills: the Code Tutor live session stops after about five paragraphs and waits, and the Grill Me design interview asks one question at a time.

Example

From the project's README, shortened and lightly edited. Before:

AI: "Here are 5 questions about your architecture:
1. Is context engineering blocking the release?
2. What's your deployment strategy?
3. Should we refactor the auth system?
4. Have you considered microservices?
5. What's your testing coverage?
Also, here's my analysis... [3 more paragraphs]"

After /breakitdown:

AI: [creates a task list tracking each question]
"First checkpoint. Is context engineering currently blocking the
release?"
 
You: "Yes, it is."
 
AI: [marks it complete, moves on]
"Got it. Second question: what's your deployment strategy?"