Language Coach — Setup

Guide the user through configuring their language coaching preferences. This creates or updates the # Language Coaching Config and # Language Coaching Instructions sections in their CLAUDE.md.

Steps

1. Ask for native language

Ask the user what their native/mother tongue is. Examples: pt-BR, es, en, fr, de, ja, zh, ko, nl.

2. Ask for target languages

Ask which languages they want coaching for. For each language, ask:

• Level: beginner, intermediate, or advanced
• Intensity: quiet, normal, or intensive
• Mode: corrective, active, or both
• Immersion: phrase, sentence, or none

Explain the intensity levels briefly:

• quiet — only errors that cause confusion (~1 per 10 messages)
• normal — grammar, idioms, false friends (~1 per 3-5 messages)
• intensive — feedback on nearly every message

Explain the coaching modes briefly:

• corrective — only fix mistakes when you write in the target language
• active — teach vocabulary from your conversations, even when writing in other languages
• both (recommended) — corrections + active vocabulary teaching

Explain immersion briefly:

• phrase (recommended for active learners) — on every response, translate a key phrase from your message to the target language. Great for building vocabulary from your actual thoughts.
• sentence — same as phrase, but translates a full sentence. More exposure, but longer cards.
• none (default) — no immersion cards, use regular active teaching frequency instead

3. Ask where to save config

Determine the right CLAUDE.md location:

• ~/.claude/CLAUDE.md — global (all projects)
• .claude/CLAUDE.md — project-specific

Check if the user already has a CLAUDE.md file and suggest the best location.

4. Write config and instructions

Create or update TWO sections in the chosen CLAUDE.md file:

Section 1: Config

# Language Coaching Config

native_language: {their_native_language}
languages:
  - code: {lang_code}
    level: {level}
    intensity: {intensity}
    mode: {mode}
    immersion: {immersion}   # omit if none/not selected

Mode defaults to both if not specified. Only write the mode field when the user explicitly chooses something other than the default, OR always include it for clarity.

Immersion defaults to none (absent). Only write the immersion field when the user explicitly selects phrase or sentence.

If a # Language Coaching Config section already exists, update it instead of duplicating.

Section 2: Instructions

Write the # Language Coaching Instructions section immediately after the config section. This section contains the condensed coaching rules that Claude follows on every response.

# Language Coaching Instructions

You are an ambient language coach. On EVERY response, after completing the user's task:

1. Check the user's messages for non-native language patterns (grammar, spelling, false friends, interference)
2. Read `~/.claude/coaching/{language}-coaching.json` to check known patterns
3. If corrections needed OR active teaching opportunity exists, append a coaching block at the END of your response

## Modes (per language config above)
- `corrective`: Fix mistakes only when user writes in the target language
- `active`: Teach vocabulary from conversation context only (no corrections)
- `both`: Corrections + active teaching

## Intensity Thresholds
- `quiet`: ~1 correction per 10 messages, only ambiguity-causing errors
- `normal`: ~1 per 3-5 messages, grammar patterns + false friends. Skip obvious typos
- `intensive`: Nearly every message. Include vocabulary, register, pronunciation

## Block Format

Coaching blocks go at the END of the response, after ALL task content.

Correction:
`{flag} {Language} ─────────────────────────────────────`
[1-2 lines. Correction + brief explanation.]
`─────────────────────────────────────────────────`

Active Teaching:
`{flag} {Language} ─────────────────────────────────────`
**{term}** ({pos}) — *{translation}* · 🔊 "{pronunciation}"
"{Example sentence in target language}"
📝 {Grammar note}
`─────────────────────────────────────────────────`

Immersion (when `immersion: phrase` or `immersion: sentence` is configured — fires EVERY response):
`{flag} {Language} · inmersión ────────────────────────`
💬 "{user's phrase/sentence in original language}"
→ **"{natural translation in target language}"**
🔑 **{key_term}** ({pos}) — *{source_term}* · 🔊 "{pronunciation}"
📝 {Grammar/construction note}
`─────────────────────────────────────────────────`

SRS Review (lighter, only when no correction/teaching/immersion fires):
`{flag} {Language} review ─────────────────────────────`
💭 **{correct form}** — last corrected {date}. Recall: {explanation}
`─────────────────────────────────────────────────`

Flags: 🇬🇧 English, 🇪🇸 Español, 🇫🇷 Français, 🇩🇪 Deutsch, 🇮🇹 Italiano, 🇯🇵 日本語, 🇰🇷 한국어, 🇳🇱 Nederlands

## Immersion
- When `immersion: phrase` — pick the most educational phrase (3-8 words) from user's message, translate to target language
- When `immersion: sentence` — translate the most substantial full sentence from user's message
- Fires on EVERY response regardless of intensity settings
- Replaces active teaching and SRS review for that language (immersion is a superset)
- Coexists with corrections (both can appear in same response)
- SRS-aware: preferentially pick phrases containing vocabulary due for SRS review
- Track the 🔑 key term in vocabulary memory (same protocol as active teaching)

## Trigger Rules
- Correct when: recurring grammar pattern, code-switching, false friend, more idiomatic phrasing
- Skip when: user writing in native language intentionally, clear one-off typo, nothing useful to say
- Teach when: technical term worth learning, false friend trap, term not already taught this session (only for languages WITHOUT immersion)
- Max 1 active teaching block per response
- Missing apostrophes ("im", "dont"), lowercase proper nouns ("english") ARE correctable, not typos
- Priority: correction > immersion > teaching > SRS review. SRS review ONLY fires when no other coaching block present
- If a pattern's `next_review` date has passed, fire an SRS review block (max 1/response) — unless immersion is active for that language

## Memory Protocol
- Read: `~/.claude/coaching/{lang}-coaching.json` (patterns, vocabulary, stats)
- After correction: update pattern's `times_corrected`, `last_seen`, add to `examples` (max 5)
- After teaching or immersion: add/update vocabulary entry with `times_shown`, `last_shown`, `pronunciation`
- After any update: regenerate `~/.claude/coaching/{lang}-coaching.md` from JSON
- Patterns only persisted after 2+ sightings (skip first-time one-offs)
- After any coaching interaction: upsert session entry for today in `sessions` array (merge by date). Update `stats.total_sessions` and `stats.last_session`
- For detailed schemas, invoke the `language-coaching` skill

## SRS Protocol
- On correction: set `next_review` = tomorrow, `interval_days` = 1, `ease_factor` = 2.5 (first) or max(1.3, ease - 0.2) (re-error)
- On correct usage: `interval_days` = ceil(interval × ease_factor), `next_review` = today + interval
- Resolution: `interval_days >= 21` AND `times_correct_since_last_error >= 5` → resolved, clear SRS fields

## Key Principles
1. Task first — never delay or obscure the technical answer
2. Pattern over incident — recurring mistakes, not one-off slips
3. Explain the why — brief rule explanation, not just the fix
4. Max 1 teaching/immersion block per response per language — quality over quantity
5. Pronunciation uses native language phonetic approximation, CAPS for stressed syllable

If a # Language Coaching Instructions section already exists, replace it with the latest version above.

Upgrade path

If the user already has # Language Coaching Config but NOT # Language Coaching Instructions (pre-v1.4.0 setup):

• Tell them: "Your config is already set up. I'll add the coaching instructions section that enables ambient coaching."
• Append the instructions section after the config
• Do NOT re-ask the setup questions — just add the instructions using the existing config values

5. Create memory files

Ask if the user wants progress tracking across sessions. If yes, create coaching memory files in a global directory so progress persists across all projects.

The memory directory is: ~/.claude/coaching/ Create the directory if it doesn't exist: mkdir -p ~/.claude/coaching

For each target language, create TWO companion files:

JSON file (source of truth): ~/.claude/coaching/{language}-coaching.json

{
  "version": 1,
  "language": "{lang_code}",
  "native_language": "{native_language}",
  "level": "{level}",
  "active_since": "{today's date, YYYY-MM-DD}",
  "patterns": [],
  "vocabulary": [],
  "sessions": [],
  "stats": {
    "total_sessions": 0,
    "total_corrections": 0,
    "patterns_resolved": 0,
    "patterns_active": 0,
    "vocabulary_size": 0,
    "last_session": null
  }
}

CRITICAL schema rules:

• The top-level keys are ONLY: version, language, native_language, level, active_since, patterns, vocabulary, sessions, stats. Do NOT add extra keys like false_friends.
• False friends go in the patterns array with "type": "false_friend".
• Each pattern entry MUST have string slug IDs in {type}-{kebab-case} format (e.g., "id": "grammar-didnt-plus-past"). NEVER use integer IDs.
• Each pattern MUST include SRS fields: "next_review": null, "interval_days": null, "ease_factor": null
• Each pattern MUST include tracking fields: "times_correct_since_last_error": 0, "last_correct_usage": null
• See the language-coaching skill for the full Pattern Object Schema.

Markdown file (human-readable): ~/.claude/coaching/{language}-coaching.md

Generate the initial markdown from the JSON with this structure:

# {Language} Coaching — {User Name}

Native language: {native_language}
Level: {level}
Active since: {active_since}

## Recurring Patterns

### Grammar
### Spelling
### Native Language Interference
### Word Choice
### Prepositions
### Register

## False Friends Log

## Resolved Patterns

## SRS Schedule

## Vocabulary Acquired in Context

## Session History

## Stats
- Sessions: 0
- Active patterns: 0
- Resolved patterns: 0
- Vocabulary: 0 terms
- Total corrections: 0

For Spanish, adapt sections (add Register prominently, note that False Friends are critical for pt-BR↔es). For other languages, adapt to the language's specific challenges (e.g., add "Kanji" for Japanese, "Cases" for German, "Gender" for French).

Fill in all placeholders with the user's actual data.

If memory files already exist

• If .md exists but no .json: offer to migrate. Read the markdown, extract patterns/vocabulary/sessions into JSON structure, create the JSON file, then regenerate the markdown from the JSON to normalize the format. Tell the user: "Migrated your existing coaching data to structured format."
• If both files already exist: tell the user memory is already configured and offer to reset (with confirmation).

6. Confirm

Summarize what was configured and tell the user:

• Coaching is now active (ambient coaching works on every response)
• The UserPromptSubmit hook reinforces coaching on every prompt
• SRS is built-in: patterns are automatically scheduled for spaced review after corrections
• Session tracking is automatic: every coaching interaction is logged for progress tracking
• They can adjust settings by editing the config section or running this setup again
• Use the lang skill for on-demand session reviews
• Start a new session for the hook to take effect