Skip to main content

02.19Concept

Cognitive chunking for AI output

AR Labs' method for presenting AI responses one piece at a time, in chunks sized to working memory, so a long answer becomes a conversation.

Break It Down on GitHub
A long answer cut into five chunks and served one at a time, sized to four working-memory slots, with a turn after each.A long answer on a tall sheet is cut by dashed lines into five chunks. One chunk, outlined in violet, sits on the table while the rest wait. Below, a turn cycle reads present, wait, respond, next. To the right four working-memory slots hold one chunk, and a crossed-out spill of five cards shows the overload being avoided.ONE LONG ANSWERWAITING1 OF 5PRESENTWAITRESPONDNEXTWORKING MEMORYABOUT FOUR CHUNKS AT ONCEFIVE QUESTIONS AT ONCEAR LABSCONCEPTSC-12ONE CHUNK AT A TIMESCALE NTS
On this page
  1. Where it comes from
  2. How it works
  3. When to use it
  4. How we apply it

Cognitive chunking for AI output is the practice of breaking a long or complex AI response into a sequence of small pieces, presenting one, waiting for the person to respond, and then moving to the next. AI assistants increasingly answer with five questions at once, a many-section analysis, or ten options to weigh. Each answer may be correct, but a person cannot hold that much in mind at once, and the result is overload rather than understanding. Chunking turns a monologue back into a dialogue. AR Labs uses it in its own agents and has published it as an open-source command for Claude Code.

Where it comes from

The idea rests on three findings from cognitive psychology.

Working memory holds about four chunks. Nelson Cowan's 2001 paper "The magical number 4 in short-term memory" revised George Miller's earlier estimate of seven plus or minus two down to about four items, plus or minus one. A response asking five questions already exceeds that.

Presentation adds load. Cognitive load theory, which John Sweller introduced in 1988 and refined in the 1990s, separates the load inherent in the material from the extra load created by how it is presented. Complex content cannot be made simple, but the extra load can be removed, which frees capacity for thinking.

Attention is a limited budget. Daniel Kahneman's Attention and Effort (1973) models attention as one limited pool of capacity shared among everything a person is doing. Several demanding streams at once draw on the same pool, so each gets less.

The same concern runs through our founding framework, cognitive ecology, which starts from the small capacity of conscious attention.

How it works

The pattern is four steps, repeated until the response is used up: chunk, present, wait, repeat.

FIG. 01
flowchart LR
  R["Long AI response pending"] --> T["Split into checkpoints and track them"]
  T --> P["Present one checkpoint"]
  P --> W["Wait for the person's reply"]
  W --> D{"Checkpoints left?"}
  D -->|"yes"| P
  D -->|"no"| E["Done"]

The AI first recognises that its pending answer is complex: several questions, a long analysis, many options, or a decision that depends on earlier decisions. It writes the parts down as a tracked list of checkpoints, so nothing is lost. It then presents only the first, waits for the answer, marks that checkpoint complete, and moves on. Because the list is tracked, the conversation can follow the person's answers without dropping the remaining items.

A request to plan a refactor, for example, would otherwise produce a full architecture analysis. Chunked, it begins with one question about how the system works today, and each later step builds on the answer to the one before.

When to use it

Chunking helps when an AI is about to deliver three or more questions at once, a multi-section analysis, a list of options to evaluate, dense reference material, or a planning session that needs a series of decisions. It is not needed for a short, single answer.

How we apply it

We released the method as Break It Down, a /breakitdown command for Claude Code published as open source under the MIT licence. Installing it is a matter of copying one command file into the Claude Code commands folder, and the pattern can be adapted to any assistant that can keep a task list.

The same principle shapes our other work. The VOX interface design answers with visual structure rather than long spoken or written blocks. And Salesbot asks its questions one stage at a time while showing its progress on screen.