fluent-speaking-import

Import the results of a Claude app voice speaking session back into Fluent — fetch the results JSON from a Slack channel, Google Drive, pasted text, or a raw transcript, validate and re-score it, conv

SKILL.md

GitHub

Speaking Practice Import (Voice Round-Trip, Step 2 of 2)

Overview

Closes the loop opened by /fluent-speaking-export: takes the results JSON produced by the claude.ai "Fluent Speaking" Project (schema defined in .claude/references/voice-project-instructions.md), validates it, converts it into one update-db.py payload, and records the session exactly like a native /fluent-speaking session — results markdown, mistakes, SM-2 reviews, streak.

This skill is the authoritative scorer. The external chat's output is evidence, not truth: validate every field, re-derive what doesn't add up.

When to Use

Trigger on /fluent-speaking-import, or when the learner's message indicates they just finished (or want to pull in) a voice session — e.g. "I'm done with voice practice", "pull in my results", "import my speaking session". Don't fire on vague mentions of speaking practice with no completion/import intent (e.g. a question about how voice mode works).

Instructions

1. Load context

python3 "${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PROJECT_DIR:-.}}/.claude/hooks/read-db.py"

Need: computed.next_session_id, computed.today, mistakes_db (existing pattern_ids), spaced_repetition (item ids), learner profile (preferences.voice_slack_channel, preferences.voice_slack_channel_id).

2. Locate the results

In order:

  1. Pasted by learner — if the message invoking the skill (or the next one) contains a fluent_voice_results JSON block, use it.
  2. Slack — if preferences.voice_slack_channel is set: load mcp__claude_ai_Slack__slack_read_channel via ToolSearch (and mcp__claude_ai_Slack__slack_search_channels only as a fallback — see below). Prefer preferences.voice_slack_channel_id if already saved on the profile — skip the channel-name search entirely and call slack_read_channel(channel_id, limit=5) directly. If no voice_slack_channel_id is saved yet, resolve it once via slack_search_channels(voice_slack_channel) and save the resulting ID to that preference field (a normal profile edit) so future imports skip the lookup. Scan the 5 newest messages (returned newest-first) for one containing fluent_voice_results. If none of the 5 match, fetch one more page of 5 older messages (10 total) and stop — do not paginate further. If still not found after those 10, fall through to Google Drive / ask the learner. If the JSON was split across a parent message and a threaded reply (the export skill does this for long cards, but results are always small — check anyway), use slack_read_thread to pull the reply and concatenate.
  3. Google Drive — load mcp__claude_ai_Google_Drive__search_files via ToolSearch; query title contains 'fluent-voice-results', take the newest match, read it with read_file_content. (The file may have been converted to a Google Doc — extract the JSON from the text representation.)
  4. Raw transcript fallback — if the learner pastes a chat transcript with no JSON block, reconstruct the data yourself: identify each question/answer exchange and score it with the fluent-speaking / voice-project-instructions.md rubric (Communication /5, Grammar /3, Vocabulary /2). Tell the learner you're re-scoring from the transcript.

If nothing is found, ask the learner to paste the JSON block from the chat.

3. Validate (authoritative-scorer pass)

  • fluent_voice_results must be true; session_ref must match an existing /results/fluent-voice-card-{session_ref}.md (warn and continue if the card is missing; abort if session_ref is absent entirely).
  • Multi-language guard: if <data_dir>/languages.json exists (multi-language mode active) and the results JSON has a language field, compare it (case-insensitive) against the active language's label. On a mismatch — e.g. the learner practiced Swedish by voice but English is currently active — stop and ask: "This session was in {results_language}, but {active_language} is your active language. Run /fluent-language switch {results_language} first, then re-import?" Importing into the wrong language's databases silently corrupts that language's streak/mastery, so never import across a mismatch without confirmation.
  • Duplicate guard: grep -l "{session_ref}" results/fluent-speaking-session-*.md — if any hit, this session was already imported. Stop and tell the learner (re-import would double-count streak/session stats).
  • Per question: each of communication (0-5), grammar (0-3), vocabulary (0-2) clamped to range; recompute score as their sum (ignore the reported score if it disagrees).
  • Per error: category must be one of grammar, formal_informal, vocabulary, spelling, prepositions, articles, missing — relabel to the closest canonical category if not; severity must be critical/moderate/minor (default moderate).
  • date: valid YYYY-MM-DD, not in the future — else use computed.today.
  • duration_minutes: integer 1-120 — else estimate 2 min per question.

4. Convert to the update-db payload

Payload field Derivation
session_id computed.next_session_id (NEVER from the JSON)
date validated results date
duration_minutes validated duration
command_used "/fluent-speaking-import"
skills_practiced ["speaking"]
skill_scores.speaking {exercises: len(questions), correct: count(communication >= 4), time_minutes: duration_minutes}
errors[] flatten questions[].errors; collapse duplicates by pattern; pattern_id = existing mistakes_db id when category+context match one, else {category}_{context_in_snake_case}; carry category, severity, your_answer, correct_answer, context
review_results[] for each card "Tutor focus" line whose pattern_id exists in spaced_repetition items: {item_id: pattern_id, quality: floor(score_of_probing_question / 2)}
new_vocabulary[] one entry per new_words[] item: item_id = snake_case content (article stripped), item_type: "vocabulary", content/answer from the entry, category: "voice_session", difficulty = learner CEFR level, initial_quality: 3, priority: "medium"
topics_covered, breakthroughs, focus_next_session pass through
session_notes "Voice session via Claude app ({session_ref}). " + results session_notes

5. Write the session results markdown

/results/fluent-speaking-session-{NNN}.md (NNN from session_id), following .claude/references/session-file-template.md exactly — same headings and parsing markers (❌, ✅, **Score:** X/10, **Accuracy:** N%, **Focus Areas:**). Set:

  • **Skill:** speaking
  • **Command:** /fluent-speaking-import
  • Add one line under the header: **Modality:** voice (Claude app) — {session_ref}
  • One ### Question N block per results question, with the learner's spoken answer, corrections, and the recomputed score.

6. Update all databases

Use the fluent-db-updater skill: pipe the Step-4 payload to

python3 "${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PROJECT_DIR:-.}}/.claude/hooks/update-db.py"

Exit 1 = fix payload and retry; exit 2 = report I/O error, nothing written.

7. Show the summary

Normal Fluent session summary: questions answered, accuracy, streak status, new words added to the review queue, focus for next session. Celebrate — this was REAL speaking practice. 🎉

Example — score recomputation

Results JSON says {"communication": 4, "grammar": 2, "vocabulary": 2, "score": 9}. 4+2+2 = 8 ≠ 9 → use 8. Review quality for a pattern probed by this question: floor(8/2) = 4.

Critical Rules

  • Never trust external data blindly. Clamp, recompute, relabel — the canonical rules live in this repo, not in the chat.
  • session_id comes from read-db.py at import time. The voice chat must never pick it — another session may have happened in between.
  • One update-db.py call, at the end — same rule as every other session.
  • Duplicate imports are blocked, not merged. update-db.py rejects a reused session_id outright (exit 1, nothing written); the Step-3 session_ref duplicate guard exists to catch it before that — re-running this skill on an already-imported session must not look like new progress.
  • Transcript fallback always exists. A voice session without a JSON block is still importable — you re-score it yourself.
  • Slack is read-only here. This skill only reads from the channel; it never posts. /fluent-speaking-export owns the one-time voice_slack_channel preference (this skill may still cache voice_slack_channel_id as a read-side optimization if export hasn't yet).
  • Auto-invocable, but only on clear completion/import intent — the duplicate-guard in Step 3 makes accidental re-firing harmless (it stops and tells the learner rather than double-importing), so err toward firing when the learner sounds done, not toward re-asking /fluent-speaking-import.