feishu english game

Use this skill for Feishu/Lark English-game sessions.

Hard rules

• Treat this as a group-chat game skill, not a generic tutoring flow.
• Keep the interaction lightweight and fun.
• Do not ask broad setup questions.
• If the mode is not specified, ask only one concise question: vocab / guess / speaking?
• If the difficulty is not specified, ask only one concise follow-up or default to cet6.
• Keep game state in Claw's working context: current mode, difficulty, target word, guess history, hint count, winner, and whether the round is open or finished.
• Do not dump JSON, raw payloads, or debug objects into Feishu/Lark.
• Reply in normal chat language.
• Use SENSEAUDIO_API_KEY from environment. Never hardcode keys.

Core modes

1. vocab mode

Use when the group wants a Wordle-like vocabulary game.

Behavior:

• Choose a target English word that fits the requested difficulty.
• Reveal only the word length at the start.
• Let users submit whole-word guesses.
• Score each guess with the deterministic helper script in scripts/english_game.py.
• Keep a visible board in chat using text or emoji. Do not rely on model-only letter scoring.
• After a correct answer or a pass, give learning value:
• ipa / pronunciation
• part of speech
• chinese meaning
• synonyms
• antonyms when useful
• one natural english example sentence

Hints may be released gradually if the group is stuck:

• chinese meaning hint
• first-letter hint
• root/prefix/suffix hint
• synonym-style hint

2. guess mode

Use when the group wants a more conversational guessing game.

Default submode is definition.

Supported prompt styles:

• definition
• example
• synonym
• antonym
• cloze

Behavior:

• Post one clue at a time.
• Let anyone in the group answer.
• Keep the pace short and chat-native.
• No large board is required.
• After the answer is solved, still provide the same learning wrap-up as vocab mode.

For v1, prefer definition unless the user explicitly asks for another submode.

3. speaking mode

Use when the group wants spoken English practice through voice messages.

Behavior:

• Post a short prompt, sentence, or question.
• Wait for a voice message.
• Download the voice file and transcribe it with scripts/asr_transcribe.py.
• Use the transcript as the basis for feedback.
• Reply with:
• what you heard
• whether it matched the target or answered the prompt
• a short correction or polish suggestion
• a better natural version when helpful

Important limitation:

• ASR transcript quality can support speaking feedback.
• ASR alone cannot produce trustworthy phoneme-level pronunciation scoring.
• Do not claim precise pronunciation grades or accent judgments unless another dedicated pronunciation model exists.
• Keep speaking feedback practical and light.

Default commands for Feishu/Lark

Use a single entrypoint concept such as /english.

Recommended command pattern:

• /english start
• /english start vocab
• /english start guess
• /english start speaking
• /english hint
• /english pass
• /english stop
• /english rank
• /english help

If the user says only 英语游戏, start mode selection.

If the user already said vocab, guess, or speaking, go directly into that mode.

Difficulty guidance

Supported labels:

• cet4
• cet6
• kaoyan
• ielts
• toefl

Selection rules:

• cet4: common core vocabulary, shorter words, familiar daily usage
• cet6: broader academic/common advanced vocabulary
• kaoyan: more abstract and formal words
• ielts: communication + academic discussion vocabulary
• toefl: academic and concept-heavy vocabulary

When uncertain, default to cet6.

Deterministic helpers

vocab scoring

Use the script below for every vocab guess so repeated letters are handled correctly:

python3 scripts/english_game.py vocab-score \
  --target-word "explanation" \
  --guess "probability"

The script prints plain text that includes:

• normalized guess
• emoji score row
• solved true/false

Use that output to update the board you show in Feishu/Lark.

speaking transcription

Use the script below for voice-message transcription:

python3 scripts/asr_transcribe.py \
  --file /path/to/audio.m4a \
  --language en

Defaults:

• model: sense-asr-pro
• response parsing: plain transcript text only
• punctuation: enabled when supported
• ITN: enabled when supported

If the user wants translation for speaking review, add --target-language zh or another supported language.

Feishu/Lark response style

Always reply in natural chat language.

Good examples:

英语游戏开始啦，选一个模式：vocab / guess / speaking

你猜对了！
单词：explanation
音标：/ˌekspləˈneɪʃən/
词性：noun
释义：解释；说明
例句：Her clear explanation helped the team understand the problem.

我听到的是：I would like to book a ticket for tomorrow morning.
整体表达很自然。把 for tomorrow morning 再连读顺一点会更像真实对话。

Bad examples:

{"text":"...","segments":[...]}

{"guess":"probability","score":["absent","present",...]}

Session management

Claw should remember within the ongoing chat:

• current mode
• chosen difficulty
• target word or target clue
• guess history
• hints already used
• who solved it
• whether a round is active

Do not ask users to repeat information that already exists in recent chat context.

Environment variables

Required:

export SENSEAUDIO_API_KEY="your_key"

Optional:

export SENSEAUDIO_BASE_URL="https://api.senseaudio.cn"

Resources

• scripts/english_game.py: deterministic vocab scoring helper
• scripts/asr_transcribe.py: SenseASR transcription helper for speaking mode
• references/integration_cn.md: Chinese interaction rules for Feishu/Lark
• references/modes_cn.md: mode-specific gameplay guidance
• references/asr_provider_notes.md: ASR defaults and capability notes