Claude Code's learning mode vs retrieval practice
The short answer
Claude Code's learning mode is the built-in Learning output style. Claude explains its choices in Insight blocks and leaves a few lines of each task for you to write, marked with a TODO(human) comment. Retrieval practice is a different method: later, with the code closed, you answer questions about what you worked on from memory, then check your answers.
The two help at different times. The Learning style has you do part of the work while you build, so you understand this change today. Retrieval practice is the method with the stronger evidence for still knowing the material a week later. If you want to keep what you build with an agent, use the Learning style on work you want to learn, and add recall practice afterwards.
What does Claude Code's learning mode do?
Claude Code calls it the Learning output style. An output style is a set of instructions that Claude Code sends with every request in a session. Anthropic's documentation describes the Learning style this way:
“In the Learning style, Claude adds the same Insight blocks as the Explanatory style and also asks you to write some of the code. Claude handles routine implementation itself. When it reaches a piece with a real design decision, such as error handling, a data structure, or business logic with more than one valid approach, it leaves a few lines for you.”
Claude marks the place with a TODO(human) comment in the file. It then sends a short request that says what is already built, what to write and what to consider, and waits. When you say you are done, Claude replies with one Insight block about your code and continues the task.
The Explanatory style is the smaller version of the same idea: Claude adds the Insight blocks but writes all of the code itself.
How do you turn on learning mode in Claude Code?
Run /output-style learning in a session, or run /config and choose Output style from the menu. Both save the choice to .claude/settings.local.json for the current project. Claude uses the new style from your next message.
To set the style by hand, add "outputStyle": "Learning" to a settings file. The value is case-sensitive: "learning" in a settings file gives you the Default style. The /output-style command ignores case. To use the Learning style in every project, set outputStyle in ~/.claude/settings.json; a project's own settings files take precedence over that value.
What are the limits of learning mode?
Five limits are worth knowing before you rely on the Learning style:
- It is an instruction, not a rule Claude Code enforces. The documentation says an output style "doesn't guarantee that something always happens or never happens."
- It applies to the main conversation. Subagents run their own system prompt, so work that a subagent does does not follow the Learning style.
- It costs time and tokens. Each task pauses for your input, and the documentation notes that the Explanatory and Learning styles produce longer responses than Default, which increases output tokens.
- Claude chooses which pieces you write. You get the parts Claude judges to be real design decisions, and those may not be the parts you most need to practice.
- It happens once, while the code is in front of you. Nothing in the style brings the material back a day or a week later.
What is retrieval practice, and what does the evidence show?
Retrieval practice means answering questions from memory instead of reading the material again. It has more than a century of research behind it.
In the reference experiment, students read short prose passages. One group read a passage once and then practiced recalling it three times. Another group read it four times. A week later, the recall group remembered 61% of the passage and the rereading group remembered 40% (Roediger & Karpicke, 2006).
The effect holds in real classrooms. A 2021 meta-analysis of 222 classroom studies, covering 48,478 students, found that quizzing raised achievement by about half a standard deviation (Yang et al., 2021).
Two limits matter for engineers. First, the size of the benefit depends on the comparison: it is large against rereading, and one meta-analysis measured it close to zero against other active methods, such as explaining the material to yourself. Second, the evidence on working professionals is thin. There are about six studies, their pooled result has a confidence interval that crosses zero, and there is no randomized trial on programming.
How do learning mode and retrieval practice compare?
They differ in when they happen and what they help with. This table compares the Learning style with retrieval practice on five points.
| Claude Code Learning style | Retrieval practice | |
|---|---|---|
| When | During the task, with the code open | Later, with the code closed |
| What you do | Write the few lines Claude leaves for you | Answer questions on the topics from memory, then check |
| What it helps with | Understanding this change, now | Recalling the material days or weeks later |
| Cost | Every task pauses for you, and responses are longer | Short sessions, separate from the task |
| Evidence | We found no published study of the Learning style itself | A century of studies on recall after a delay; few on professionals, none randomized on programming |
Should you use learning mode or retrieval practice?
Use both, for different jobs. Turn on the Learning style for work in a library, language or part of the codebase you want to learn, and leave it off for routine tasks where the pauses only slow you down. Then practice recall on the topics that session covered, a day or more later.
An Anthropic trial of 52 developers learning a new Python library supports writing some of the code yourself. The developers who had an AI chat assistant scored about 15 percentage points lower on a quiz taken minutes after the task than developers who used documentation and web search (Shen & Tamkin, 2026). Inside the AI group, the developers who asked conceptual questions, asked for explanations with the code, or questioned the code after generating it scored 65-86%. The developers who delegated the task, relied more on the assistant as the session went on, or pasted errors back until they went away scored 24-39%. Each of those groups had two to seven people, and the study did not assign anyone to a usage pattern, so treat the split as a pattern, not proof. The trial also used a chat assistant, not Claude Code.
The Learning style asks you to work the way the higher-scoring developers worked. That trial measured understanding minutes after the task, though, not a week later. Recall after a delay is what the retrieval-practice studies measure, and it is the part the Learning style does not cover.
Where Atomic Reps fits
Atomic Reps adds the recall step to the tools you already use. When your coding agent finishes a task in Claude Code, GitHub Copilot, Cursor or Codex, it asks you questions on the topics that task worked on, and you answer each one with a letter, from memory.
It does not replace the Learning style. The Learning style has you write part of the task; Atomic Reps has you recall the topics behind it.
See how it works in your coding agent