cefr language tutor prompts

You are likely seeing conversational language tutors that greet users on every turn, repeat broken grammar as if it were correct, or default to English even for C2 learners. These are not personality flaws; they are state management and instruction capping failures that result in "schoolmasterly" agents that users quickly abandon.

This skill provides a battle-tested framework for building adaptive CEFR-based language tutors. It moves beyond generic prompting by treating conversation flow as a state problem, ensuring instructions change based on turn count and learner proficiency. It focuses on the specific nuances of voice-first tutoring, where unspoken text like phonetics or markdown formatting can break the learner's immersion.

What it does

• Level-Capped Immersion enforces target-language regimes based on CEFR levels (A1-C2) to prevent advanced learners from getting stuck in "native language first" modes.
• State-Aware Greetings provides logic to prevent repetitive "Hi [Name]" openers by moving turn-sensitive instructions out of the static rule block.
• Natural Error Correction handles speech-to-text slips by offering repairs as questions ("Did you mean...?") rather than decreeing grammar rules.
• TTS-Safe Output strips markdown, bolding, and invented phonetics that cause speech synthesizers to hallucinate or stutter.
• Instructional Guardrails implements a "two-ask" limit for corrections and prioritizes warmth over rigid rule adherence to prevent learner fatigue.

How it works

1. Analyze the learner profile to determine if the CEFR level should override the user's selected immersion mode.
2. Inject state-specific instructions into the per-turn prompt (e.g., "Not the first turn") to control greeting behavior.
3. Execute the repair logic when a native speaker would not say the input sentence, framing it as a plausibility check.
4. Clean the output string for voice synthesis by removing all non-spoken formatting and phonetic guides.

Frameworks & tools

Works with any LLM (Claude, GPT-4o) and TTS engines (ElevenLabs, OpenAI Voice). Best implemented in a backend that tracks turn state and user metadata.

Why this beats prompting it yourself

Most prompts fail because they include rules that contradict the learner's actual level or repeat static instructions that only apply to turn one. This skill provides the "non-obvious" logic for state management and immersion capping that prevents the tutor from sounding like a script.

Use cases

• Building a voice-first language app where formatting like **word** breaks the audio.
• Fixing a tutor that ignores the "C1" level and speaks 80% English to the user.
• Implementing a correction flow that feels like a conversation rather than a red-pen audit.

Known limitations

Does not include curriculum design or placement testing logic. Requires a backend capable of tracking turn count to implement the greeting logic effectively.