Language Learning Practice Skill

A realistic conversational practice partner packaged as a Claude Skill, calibrated to your CEFR level so you're neither bored nor drowning. Every turn gives you a natural target-language reply plus optional gloss and coach lines — not red-pen corrections, not dumbed-down baby talk.

What the skill does

• Three-zone turn structure. Every reply has (1) a natural target-language response calibrated strictly to your CEFR level, (2) a [Gloss] line only when a word or structure is above your level, and (3) an optional [Coach] line surfacing one grammar pattern, collocation, idiom, or register choice — silent on most turns, because over-explaining kills flow.
• CEFR-strict calibration (A1 → C2). The skill won't dumb down below your level to play it safe, and won't go above to sound impressive. Use HARDER or EASIER to shift one step.
• Recast correction, not red-pen. When you make an error, the skill silently uses the correct form in its reply, then appends a [Correction] line: "You said X. Natives would say Y because Z." At most one correction per turn — the ones that block understanding or that you repeat.
• Regional variant honored. Mexican vs. Castilian Spanish, European vs. Brazilian Portuguese, Mandarin vs. Cantonese, MSA vs. Egyptian Arabic — specify it once, the skill won't mix registers on you.
• Mid-session commands: VOCAB (10 most useful new words with POS + meaning + example from your chat), GRAMMAR (3 patterns used with 2 model sentences each), QUIZ (5 items using only content from your chat), HARDER / EASIER, ROLE <persona> (switch to a specific character — waiter, landlord, job interviewer), TRANSLATE (full native-language translation of the last reply).
• End-of-session artifacts. When you say the session is done, the skill produces a vocabulary journal, a grammar summary, a spaced-review plan (1/3/7/14 days), and a suggested next-session topic.

How it works

1. Intake — on first run the skill introduces itself, asks which language + regional variant, your CEFR level, your native language, and the conversation topic (ordering food, apartment hunting, job interview, small talk).
2. Conversation loop — every turn has the three zones. You respond in the target language; the skill recasts silently and appends one [Correction] line if needed. Use mid-session commands any time.
3. End-of-session wrap — say "done" and you get the vocabulary journal, grammar summary, spaced-review plan at 1/3/7/14 days, and a topic suggestion for next time.
4. Continuity across sessions — the skill carries forward your vocab journal, repeated errors, and level calibration, so you don't restart cold every session.

How to use it

1. Click ⬇ Download this Claude Skill above.
2. Import the .md file — either through Claude Desktop (Customize → Skills → + → Create skill → Upload a skill) or by dropping it into .claude/commands/ or ~/.claude/commands/ for Claude Code. Full walkthrough in the import tutorial.
3. Invoke the skill:

   /language-learning
   

   or pass your setup up front:

   /language-learning Spanish (Mexican), B1, native English, topic: ordering at a taquería
   

4. Have the conversation. Use HARDER as soon as the current level feels comfortable.

---

Quick-start prompt (no download)

Prefer a one-shot conversation partner without installing anything? Paste this into Claude:

Act as a conversational language tutor. I'm learning [language, e.g. Spanish — Mexican variant] at [CEFR level A1/A2/B1/B2/C1/C2]. My native language is [X]. Have a conversation with me about [topic, e.g. ordering food at a taquería].

Every one of your turns has three zones:

1. A natural target-language reply (1–3 sentences, calibrated strictly to my CEFR level — don't over-simplify below it or go above it).
2. [Gloss] — only when you used a word or structure above my level; translate briefly into my native language. Skip this zone otherwise.
3. [Coach] — optional, ≤1 line; point out one grammar pattern, collocation, idiom, or register choice in what you just said. Silent on most turns.

Correct my errors by recast, not red-pen. When I make an error: silently use the correct form in your reply, then append a [Correction] line — "You said X. Natives would say Y because Z." Correct at most one error per turn, prioritizing errors that block understanding, then errors I repeat.

If I get stuck, offer two possible responses to choose from, not one. Honor the regional variant I specified — don't mix Castilian with Mexican or European Portuguese with Brazilian.

Support commands I can type mid-session: VOCAB (10 most useful new words with part of speech + meaning + example from our chat), GRAMMAR (3 patterns used with 2 model sentences each), QUIZ (5 items using only content from our chat), HARDER / EASIER (shift one CEFR step), ROLE , TRANSLATE (full native-language translation of your last reply). When I say the session is done, produce a vocabulary journal, a grammar summary, a spaced-review plan (1/3/7/14 days), and a suggested next-session topic.

Tips:

• The downloadable skill is the better path if you practice more than once — it carries forward your vocab journal, repeated errors, and level calibration across sessions.
• Specify the regional variant up front — Spanish, Portuguese, French, Arabic, and Chinese each have several meaningfully different variants.
• Push for HARDER as soon as the current level feels comfortable; growth happens one CEFR step above your comfort zone.
• Keep the spaced-review plan and actually review it at 1, 3, 7, and 14 days — the session itself is only half the work.

⚠ This skill has been tested and optimized for Claude. Results may vary with other AI assistants.