~/jamie/projects/spelling-ladder
● live · PROJECT · Nov 2026 · built with Claude (Desktop, then Claude Code)

Spelling Ladder

My daughters wanted spelling practice that grows with them. So we built one: hear a word, type it, climb from 1st grade to 12th.

→ try it live
┌─ fig. 1 ─ listen660 × 490 ─┐
One button, one box. Every word is read aloud. The sentence fills in the blank so a homophone is never a guess.

┌─ HOW IT WORKS

  • 01Hear it, type it. A big speaker button replays the word, and they can hear it used in a sentence.
  • 02514 real words, grades 1–12. Grades 1–5 from the Basic Spelling Vocabulary List (Graham, Harris & Loynachan); 6–12 are hard-to-spell words placed by grade with the VXGL list, plus classic “spelling demons” like necessary, rhythm and February.
  • 03Climbing. Move up a grade with 18 of your last 20 right, or 10 in a row.
  • 04Review that remembers. About 30% of questions come from earlier grades, weighted toward words you've missed. A missed word keeps coming back until it's spelled right twice.
┌─ BY THE NUMBERS ───────────────────────────────────┐
│ words          514    grades       1 → 12          │
│ audio clips    1,028  per word     word + sentence │
│ level up       18/20  or           10 in a row     │
│ review share   ~30%   first        missed words    │
│ accounts       0      server code  0               │
└────────────────────────────────────────────────────┘
┌─ fig. 2 ─ the miss─┐
Show the fix, not the failure. The right spelling as tiles, with the letters to fix lit up, and what the word means.
┌─ fig. 3 ─ start─┐
A profile per kid. Progress is saved in the browser. No accounts, nothing to sign into.

┌─ HOW WE BUILT IT

  1. 01Simple on purpose. It started as one HTML page and a word list. It's since been rebuilt in Next.js, but it still ships as plain static files on Cloudflare Workers — no server, no accounts.
  2. 02Real lists, checked. Published graded word lists instead of invented words. Each word carries a sentence, a short meaning and three plausible misspellings.
  3. 03Natural audio, made once. All 1,028 clips were generated on my laptop with Kokoro-82M. A word read alone sounded robotic, so the model says “The next word is: necessary.” and the app keeps only the last word, using the model's word timings.
  4. 04Shaped by the kids. The first version showed too much at once. We cut it to one big button, one text box and friendly feedback. A multiple-choice mode exists but stays off.
  5. 05Ship every change. Commit, push, deploy — because the real test is my daughters on the live site.
the possum, in thread

┌─ WHAT I LEARNED

[YOUR TAKEAWAY — e.g. when a model sounds wrong in isolation, give it the context it was trained on and cut the part you need.]