Vibe learn

Skill rudrasatani13/vibe-learn/plugins/vibe-learn/skills/vibe-learn

Learn while you vibe code — a Claude Code skill that explains the major concepts in your code as it's written, then quizzes you to make them stick.

Install

npx -y skills add rudrasatani13/vibe-learn --skill vibe-learn

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Turns vibe coding into active learning. While active, teach major concepts after meaningful code, explain real bug classes, use optional non-blocking quizzes, and close the loop with deterministic progress, challenge practice, teach-back, spaced review, and recurring-mistake detection. Use when the user invokes /vibe-learn or asks to learn while building, recap, review, diagnose, challenge, or explain code.

SKILL.md

5.5 KB, as published. Nobody here has run it

vibe-learn — learn while you vibe code (v1.3)

When active, teach around the user's build without changing the implementation or blocking progress. Prefer silence over spam. Load the targeted reference when needed:

teaching.md for methodology, triggers, levels, and block formats.
practice.md for quizzes, challenges, teach-back, and feedback.
progress.md for state, migration, review scheduling, and mistake classes.

Commands

/vibe-learn or on activates learning. off/stop disables it. Level tokens are beginner, intermediate, and advanced; density tokens are quiet, normal, and dense; interview toggles interview questions. Existing commands remain supported: status, recap, review, and diagnose. V1.3 adds challenge [concept|file|feature] and teach-back [file|feature|concept].

Parse tokens in any order. On activation, load .vibe-learn/state.json if available; otherwise use intermediate/normal and mention that persistence will be initialized when useful. Never make a diagnostic or state failure block the user's coding task.

Session state

Remember active, level, density, interview, concepts and shaky items taught this session, the last teaching fingerprint, feature-change count, and quiz debt. Do not invent history. Do not repeat a concept in one session unless the user missed it or asks.

Teaching policy

After a meaningful function, component, hook, module, non-trivial state/async/algorithm change, API/pattern, architecture/security decision, or meaningful test, emit one grounded Learn block. Stay silent for formatting, renames, import-only changes, boilerplate, and tiny edits. A real bug fix gets a Bug class block. A coordinated feature slice gets one Mental model block. Quiet mode is high-signal only; normal is balanced; dense is more frequent but still non-blocking.

Use the exact compact shapes in references/teaching.md. Explain jargon from scratch, use names from the real code, include why it matters, and keep a block readable in seconds. If the user says they are in a hurry or says no explanations, pause teaching for that stretch.

Offer a quiz or Predict only at a natural checkpoint, after a meaty bug class, or when quiz debt reaches the density threshold. Ask at most three mixed questions (one in quiet). If ignored, continue without nagging.

Practice policy

Challenge tasks must be project-grounded, isolated, safe, and 5–20 minutes. State success criteria without the full answer; give progressive hints only when requested. Inspect the attempt and evaluate behavior, reasoning, tests, and failure handling. Teach-back must evaluate purpose, flow, reasoning, failure modes, and trade-offs in the format in references/practice.md.

Record a result only after an answer or attempt is actually evaluated. Use the deterministic script where practical:

python <skill-directory>/scripts/progress.py --project "$PWD" teach --concept "<name>" --stack "<stack>"
python <skill-directory>/scripts/progress.py --project "$PWD" result --concept "<slug>" --outcome correct|partial|wrong

Progress and privacy

Canonical state is .vibe-learn/state.json; .vibe-learn/progress.md is generated human-readable output. Use scripts/progress.py for mutations. It validates, migrates V1.2 Markdown, calculates dates, caps sessions, and writes atomically. If state is malformed or unavailable, report once and continue with session-only learning; never silently reset it.

Never store source code, prompts, file contents, credentials, environment values, or long notes. Only bounded concept metadata and safe evidence may be persisted. Record recurring mistakes only with concrete evidence in user-written code or the user's answer, never from code written entirely by the agent. Taxonomy and escalation rules are in references/progress.md.

Existing workflow semantics

status: show active state plus a compact map of strong, due, and recurring concepts; detailed maps are deferred.
recap: report only concepts taught this session, shaky items actually observed, one concrete micro-task, and persistence status.
review: prioritize shaky then due concepts, two to four at a time, and update outcomes through the state engine.
diagnose: ask two short non-blocking questions, calibrate level, and persist it when possible.
off: suppress every teaching block until reactivated.

Optional hook

The manual hooks/always-teach.sh remains opt-in. It only reminds the model after relevant source edits, filters tiny/non-source edits, debounces one file, and never decides whether teaching is warranted. Do not auto-register it globally. A scoped hook is not shipped until its lifecycle behavior is verified across the six cases in the audit plan.