Spellsketch
About Spellsketch
Five questions, answered honestly.
In August 2026, the US Department of Education said every education technology product should be clear about five things. Here are our answers, including the parts we can’t answer well yet.
01What problem does it solve?
Many people know the words but can’t put together one sentence about a picture, and when they do, they can’t tell what they left out.
Here, you talk about a picture in English, and the picture grows only with what you say: no color yet means checkered , and if you don’t say where the cat is, it won’t be under the table. You can see what’s missing for yourself, without waiting for someone to mark it. It covers basic English: the names of things, numbers, colors, positions and a few set sentence patterns.
02When should you use it?
Once a day for ten minutes, as speaking practice alongside picture books or classes.
It isn’t the main course, and it doesn’t replace a teacher, picture books or a textbook. If you don’t know the alphabet and some basic words yet, it’s too early: start with picture books, songs or a beginner class, and come back once you know some words.
03Who is it for, who isn’t it for, and what doesn’t it teach?
It’s for children learning basic English, with a parent nearby. Adult beginners can use it too.
It isn’t for people who already speak fluently, or for anyone preparing for an exam or looking for grammar lessons. It doesn’t teach phonics, reading, grammar or conversation. It reads words and sentences aloud for you to repeat, but it doesn’t score pronunciation.
04How long does it last?
The starter pack takes about three weeks, the word packs at launch about three months, and then there’s a new pack each month.
Words run out, and we won’t pretend this app will keep you busy for years. After three months, use the evidence in the next answer to decide whether to keep going.
05What’s the evidence?
There’s no evidence yet that this app improves anyone’s English. We’ll measure it ourselves and show you whatever we find.
How we measure
On every picture you see for the first time, we count how many details your first try gets right (things, colors, numbers, positions); retries and pictures you’ve seen before don’t count. Week by week, that makes a curve. It’s on each member’s progress page in the app, for free.
When progress stalls
The app adjusts on its own first, and tells you each time what it changed and why. You can always change it back.
| Order | What we see | What the app does first |
|---|---|---|
| Any time | The app often can’t make out what you say | Switches to typing, or to building sentences by tapping words |
| First | The curve isn’t rising, or the same kind of detail keeps getting missed | More pictures for that kind of detail, one fewer thing per picture, one-word hints on, and each day starts with listening before speaking |
| Then | Still flat | Pauses new words to practice the known ones; invites family to describe a picture together |
| Last | Several changes tried, enough days practiced, still no rise | Stops adjusting and tells you honestly: this approach may not suit you right now. We suggest another way, such as a human teacher or picture books, or taking a break. The cancel button is right on that page: no pleading, no discounts. |
It adjusts the other way too: when progress is fast, the pictures get harder and new words open sooner. If there hasn’t been enough practice, we’ll only say “not enough practice to tell yet,” never that it doesn’t work. When each step starts depends on your own numbers, which are on the progress page in the app; we’ll add them here once we have real numbers.
Research behind it, not evidence
The approach uses four methods with a research base: describing pictures, saying things yourself (output practice), pulling words from memory (retrieval practice) and practicing again just before you forget (spaced review). That research is about the methods, not this app. Whether this app works is what the curve above will show.
06What the AI does, and where the data goes
The AI only judges which details of the picture your sentence got right. It doesn’t write sentences, and it doesn’t draw.
The judge is Claude Haiku, from Anthropic. It reads a description of the picture, compares it with your sentence and judges each detail: said, not said, or said wrong. On a test set of over two hundred sentences, it judged about nine in ten whole sentences correctly. When it gets one wrong, tap “Report a mistake” on the result; we read the reports and fix what’s wrong. If we change models, this page changes too.
Your sentence goes to our servers first. We pass only the sentence and the picture’s details on, to Anthropic to judge and to OpenAI to check the wording, without names or account details. Every picture is drawn and reviewed in advance; nothing is generated during play. We don’t collect real names or photos, and we don’t keep voice recordings. How long we keep data is set out in our privacy policy.