learnwy/english-learner

English Learner

Personal vocabulary assistant with persistent storage, mastery tracking, and spaced repetition.

Three auto-intercept modes (passive, hook-driven)

| Mode | Trigger | Action | |---|---|---| | English intercept | User writes in English (EN ratio ≥ 0.6) | If issues found: show 1–3 corrections in a "💡 English Tip" table. If clean: render exactly "✅ English looks fluent — no issues found." Always call vocab record-input with had_issues + issue_count. | | Chinese intercept | User writes in Chinese (CN ratio ≥ 0.3, OR contains 翻译/怎么说/用英语 intent). Always fires — never silently skipped. Tiered output: full mode for normal messages; light mode when ≥ 500 chars OR contains code keywords (代码/编程/bug/修复/重构/编译/部署/配置文件). | Full: "🌐 中译英" block — corrections (if any) + 2–3 EN translations + 2–3 vocab + batch_save. Light: "🌐 中译英 (light)" — 1 EN translation + optional 1 vocab. Both call record-input with language: zh. | | Other-language intercept | User writes prose in JA / KO / ES / FR / DE / RU / AR / etc. (low EN, low ZH ratios). Always fires. Tiered: full mode for normal, light mode when ≥ 800 chars. | Full: "🌐 Translate & Learn" — detect language, translate, 2–3 vocab, auto-save. Light: 1 EN translation + optional 1 vocab. Both call record-input with language: ja/ko/other. |

All modes wired through learnwy-dispatch (UserPromptSubmit + Stop hooks). They never block the user's actual task — corrections come first, work continues.

Every fired turn logs an entry to prose_log so we can compute fluency rate, by-language activity, and 30-day trend (vocab prose-stats).

Detailed rules, exclusion filters, and Markdown formats: see references/intercept-modes.md.

Active modes (explicit user request)

| Trigger | Mode | What happens | |---|---|---| | Single word / phrase / sentence | Lookup | vocab batch_get → AI generates if new → batch_save → render card | | 学习 / review / quiz | Quiz | quiz generate 5 all low_mastery → ask user with AskUserQuestion per item | | stats / 统计 | Stats | vocab stats → render summary | | 查单词 / 学英语 / 翻译请求 | Lookup | Same as single-word path |

Response formats (word card, phrase card, sentence breakdown, quiz prompt, stats): see references/learning-mode.md.

When NOT to fire

The intercept skips entirely only for inputs that aren't prose at all:

• Pure code / shell commands / file-path-only lines (matched by looksLikeNonProse)
• Single-line Use Skill: directives
• Anything shorter than 4 chars (or 2 chars if it contains CJK)

Tech-heavy or long Chinese is NOT skipped — it falls into light mode instead (1-sentence translation, optional 1 vocab). See [feedback-learning-tools-fire-everywhere] in memory: silent gates hurt the long-term dataset, so we always fire, just adjust verbosity.

English fluency check is special: if no issues, render the one-line ✅ English looks fluent ack — that is the fire, not a skip.

Prerequisites

• Node.js ≥ 24 (uses built-in node:sqlite)
• Writable home directory: data lives at ~/.learnwy/english-learner/data.db

Subcommands

Single CLI entry scripts/cli.cjs dispatches everything:

# vocab — CRUD + batch ops
node scripts/cli.cjs vocab batch_get '["word1","word2"]'
node scripts/cli.cjs vocab batch_save '[{"word":"...","definition":"...","phonetic":"...","examples":[]}]'
node scripts/cli.cjs vocab get_word|save_word|get_phrase|save_phrase|log_query|stats|update_mastery ...

# corrections — persist usage tips from the English intercept
node scripts/cli.cjs vocab record-correction '[{"original":"imrpove","corrected":"improve","reason":"typo","words":[{"word":"improve","definition":"改进；改善","phonetic":"/ɪmˈpruːv/"}]}]'
node scripts/cli.cjs vocab top-corrections [limit=5]
node scripts/cli.cjs vocab corrections-stats

# prose log — log every prose input (clean or with issues), then aggregate fluency
node scripts/cli.cjs vocab record-input '{"language":"en","text":"<original>","had_issues":false,"issue_count":0}'
node scripts/cli.cjs vocab prose-stats          # totals, clean-rate, by-language, 30-day window
node scripts/cli.cjs vocab recent-prose [limit=20]

# sentence — input classification + word extraction
node scripts/cli.cjs sentence classify <text>
node scripts/cli.cjs sentence parse <sentence>
node scripts/cli.cjs sentence batch_check <words>

# quiz — review session
node scripts/cli.cjs quiz generate [count] [type] [focus]
node scripts/cli.cjs quiz review [limit]
node scripts/cli.cjs quiz summary

# spaced repetition + linking
node scripts/cli.cjs link-wiki        # cross-link with llm-wiki

# report — generate a self-contained static HTML dashboard from the SQLite DB
#   sections: Overview · Due now · Activity · Words · Phrases · Corrections · Fluency · Materials
#   Fluency section shows overall + 30-day clean rate, by-language breakdown, and 20 most-recent inputs.
node scripts/cli.cjs report                       # writes ~/.learnwy/english-learner/report.html
node scripts/cli.cjs report --output <path>       # custom destination
node scripts/cli.cjs report --json                # also dump intermediate report.html.json
node scripts/cli.cjs report --open                # open in default browser after generation

# install/uninstall hooks (registered with .claude/ + .trae/)
node scripts/cli.cjs install   --scope global --target both
node scripts/cli.cjs uninstall --scope global --target both

vocab batch_* are mandatory whenever ≥2 items are involved.

Storage

~/.learnwy/english-learner/
├── data.db              # SQLite — words, phrases, history (schema in references/data-storage.md)
└── memory/
    ├── SOUL.md
    └── USER.md

Schema is migrated automatically by src/shared/db.ts (versioned migrations). Spaced-repetition columns (next_review_at, interval_days) drive the SessionStart push of 3 due cards per day.

Hooks

Wired through learnwy-dispatch (single Node process — no per-skill spawn):

| Event | Lib function | Purpose | |---|---|---| | UserPromptSubmit | lib/prompt-scan.ts | Inject English/Chinese intercept reminder | | Stop | lib/stop-scan.ts | After each assistant response, surface 2–4 advanced words found in the response (asks user; never auto-saves) | | SessionStart | lib/session-scan.ts | Once per day, push 3 due-for-review cards (spaced-repetition algorithm) |

Each scanner is a pure (payload | message) => string | null. Side effects (DB writes) live behind lib/*-store.ts.

Boundaries

This skill only: vocabulary capture, lookup, mastery, spaced repetition, intercept-driven correction.

This skill does not: full essay editing, grammar tutoring beyond the 3-issue cap, native-language tutoring beyond Chinese, or persistent memory of preferences (use learnwy-status for cross-skill stats).

See also

• references/intercept-modes.md — detailed English + Chinese intercept rules and formats
• references/learning-mode.md — quiz / lookup / stats response formats
• references/data-storage.md — SQLite schema + record JSON shapes