Lucia's Dictionary · Classroom Relay

Inspiration

Chinese-speaking parents often see English classroom instructions, reading prompts, and science sentences that their children bring home, but may not feel confident explaining them. Translation tools provide a one-time meaning, flashcard apps remove the classroom context, and homework assistants often jump directly to the answer.

I built Lucia's Dictionary to close this classroom-to-home learning gap. The goal is not to do a child's homework, but to turn one real classroom sentence into something the child can understand, practice, and remember with a parent.

What it does

Lucia's Dictionary turns a real classroom sentence into a private, personalized micro-lesson. A parent or child can type, paste, photograph, or upload an English sentence; the app creates readable word cards with Chinese meanings, phonetics, learning bands, and pronunciation.

During OpenAI Build Week, I added Classroom Relay, a complete learning flow that selects up to five priority words from local review state, explains every choice ("new word," "due today," "unsure last time," or "forgot last time"), practices listening, meaning recall, and an original-sentence cloze, updates spaced review, remembers real classroom encounters, and gives the parent one immediate follow-up prompt.

No account or child profile is required. Learning history, source sentences, recommendations, and mission results remain in the current browser.

What makes it different

Most vocabulary tools teach isolated words. Classroom Relay preserves where the word came from and carries that context through recommendation, practice, review, and parent follow-up.

Its personalization is explainable: families can see why every word was selected instead of trusting an invisible ranking system. GPT-5.6 and Codex helped build the product, while the child-facing learning loop is intentionally local and deterministic to avoid uploading children's learning history.

What was added during Build Week

Lucia's Dictionary existed before the July 13 submission period. The previous version supported sentence input, optional OCR, word cards, pronunciation, a local wordbook, basic spaced review, and quizzes.

The new Build Week work includes the Classroom Relay mission model, an explainable word-priority engine, listening/meaning/original-sentence cloze activities, multi-sentence encounter memory, a completion summary and parent handoff prompt, backward-compatible wordbook migration, and automated coverage for the complete mobile learning flow.

How I used Codex and GPT-5.6

I used Codex with GPT-5.6 as an engineering and product-design collaborator throughout the Build Week extension. It helped me inspect the existing architecture, identify the missing classroom-to-home learning loop, design the explainable recommendation and persistence models, implement the vertical slice, safely migrate wordbook data, add automated tests, and verify the mobile production experience.

I made the final product and engineering decisions: do not generate homework answers, do not create child accounts, do not send learning events to analytics, and make every personalized recommendation understandable to parents. The dated commit history and Codex Session ID 019f675a-a32d-7eb0-9e13-4d4df7cd9969 document this work.

Challenges we ran into

The main challenge was adding meaningful personalization without weakening privacy and offline constraints. A cloud-generated lesson would introduce child-content transmission, latency, cost, and unpredictable output. I instead built a local ranking engine and used the child's original classroom sentence as the learning context.

I also needed to extend an existing wordbook without breaking exported data. The new model migrates legacy source sentences into bounded encounter history while preserving mastery and review scheduling.

Accomplishments that I'm proud of

• Built a complete classroom-to-home learning loop, not an isolated feature.
• Made every personalized recommendation explainable.
• Preserved the original classroom context through practice and review.
• Added no child account or new learning-data network flow.
• Passed 97 unit tests, 6 Cloudflare runtime tests, and 7 mobile end-to-end tests.

What I learned

Personalization becomes more useful when families can understand why a learning decision was made. I also learned that Codex was most valuable not simply for generating code, but for connecting the project's existing OCR, dictionary, speech, review, offline, and testing systems into one coherent product experience.

What's next

Next, I want to evaluate whether original-sentence practice improves later recall and whether children complete more due reviews. Possible extensions include spelling and dictation, better confusion-based distractors, and a seven-day parent summary. Account sync, generated explanations, and speech scoring will only be considered after a separate child-privacy and content-safety review.

Built With

• codex
• gpt-5.6