lesson-kit

Correct your AI once. Claude Code notes the correction and applies it in your next sessions.

You fix a mistake in one session. You open a fresh session and the AI makes the same mistake again, because it starts from zero every time. The usual fix is writing the rule into a CLAUDE.md yourself, and that only happens if you remember and have the time.

lesson-kit does the noting for you:

1. You correct the AI the way you normally would.
2. It saves the correction as a short rule and tells you in one line: Learned [ui, global]: ...
3. At the start of every later session, the rules that apply are loaded back in.

you:     bukan button, harusnya native segmented control
claude:  (fixes it) Learned [ui, global]: Use a native segmented control for switching views, not custom buttons.

Status: early but real. Capture, save, and reload work today. It has not been used long enough to know how it behaves after months of lessons.

It writes instructions that load into every future session. A wrong lesson follows you around until you remove it. Look at what it saved from time to time with /lesson-kit:lessons, and delete anything you disagree with.

It does not make the model learn. No training happens. It is notes that get read again at the start of each session.

Capture is not guaranteed. The AI itself decides whether your message was a real correction. And it is only asked to decide when your message contains a correction word, such as no, wrong instead should be bukan salah harusnya. The full list is CORRECTION_MARKERS in scripts/lesson_kit.py. A correction with none of those words is not captured.

Lessons have limits. At most 30 are kept, ranked by how often they came up and how recently. A lesson not seen for 120 days expires. Two lessons that say nearly the same thing are merged.

Scope matters. A lesson is either global (loads everywhere) or project (loads only in that git repository). The AI picks the scope when it saves. If a lesson lands in the wrong one, remove it and correct the AI again with a clearer statement, for example "in every project".

It costs a little context. Saved lessons are added to each new session. The instruction to save a lesson is added only to messages that contain a correction word.

macOS and Linux only. The store uses file locking that Windows does not have.

One Claude Code version is verified. Claude Code 2.1.278. Hooks are a young feature, so a different version may behave differently.

Claude Code, in any of its forms: the terminal CLI, the Code tab in the Claude desktop app, or the VS Code and JetBrains extensions. It needs hooks, which run on your machine.

Nothing else to install.

/plugin marketplace add Adithya-13/lesson-kit
/plugin install lesson-kit@lesson-kit

Start a new session so the hooks are picked up.

Work as you normally do. When the AI gets something wrong, correct it in plain words. When it saves a lesson you will see one line, Learned [kind, scope]: ....

To see or remove what it saved:

/lesson-kit:lessons

One file: ~/.claude/lesson-kit/lessons.json. It is plain JSON. Set LESSON_KIT_DIR to keep it somewhere else. It is never sent anywhere except into your own AI conversation as context.

/plugin uninstall lesson-kit@lesson-kit

Then delete ~/.claude/lesson-kit if you also want the saved lessons gone.

Two hooks, one script, one file.

• UserPromptSubmit: if your message contains a correction word, it adds a short instruction asking the AI to decide whether to save a lesson, and how.
• SessionStart: it loads the lessons that apply, global ones plus the ones for the current git repository.
• scripts/lesson_kit.py add | list | remove: the AI calls add to save. /lesson-kit:lessons calls list and remove.

• Lessons stay in one local file on your machine.
• Nothing here contacts any server. Saved lessons reach the AI only as conversation context, like anything else in a session.
• A lesson is only created from something you wrote in your own message, never from file contents or tool output.

pip install pytest
pytest tests -q

MIT. See LICENSE.