Assignment Decoder

Assignment Decoder turns confusing homework into a plan any learner can start: bite-size steps, read-aloud, picture maps and simple words in their home language. Free, offline, no login.

Comment

Inspiration

Homework doesn't come with instructions for the instructions. A prompt like "Substantiate your thesis with a minimum of three primary sources, formatted in MLA" is clear to a teacher. To an English learner, a student with ADHD, a kid reading on a cracked phone, or a parent trying to help, it can feel like a wall.

The hackathon's line, "Learning has no single doorway," gave me the shape of the project: take the same assignment and open it through more than one door, across language, ability, connectivity and cost.

What it does

A student picks their grade level (preschool to college) and home language, pastes an assignment, and gets a plan they can actually start. They choose the format that works for them right now:

  • Visual view. It changes with age and task: a picture road for young kids, an experiment view with a live graph for labs, an argument blueprint with a word budget for essays, and a synthesis matrix for literature reviews.
  • Bite-size steps. A checklist with time estimates, plus a Focus mode that shows one step at a time.
  • Listen. The plan read aloud with the device's built-in voice.
  • Simple words. The assignment in plain language, with tricky words explained and translated into 10 home languages.

It also writes a note for family in their own language, gives a quick check that the student understood the task (not the subject), and offers an easy-read font, high contrast and adjustable text size.

It explains the assignment. It never does the assignment.

How we built it

  • Frontend: plain HTML, CSS and JavaScript with no framework, so it loads fast on old devices and slow connections.
  • AI: one serverless function sends the assignment, grade level and language to Claude (Anthropic's API) and asks for a plan in a fixed JSON shape using structured outputs. The prompt adapts the number of steps, vocabulary and amount of help to each grade level, and forbids writing any part of the student's work.
  • Offline: a service worker keeps the app working after the first visit, and the last plan is saved on the device. There's also a printable one-page checklist.
  • Cost: a teacher can share a link with the whole plan compressed inside the URL, so students who open it need no account, no server and no AI call. Read-aloud uses the browser's built-in speech, which is free. A live decode costs roughly

$$ \frac{3{,}000 \text{ input tokens} \times 4 + 2{,}500 \text{ output tokens} \times 20}{10^6} \approx 0.06 \text{ USD} $$

  • Hosting: Vercel, with the code on GitHub.

I built it with Claude Code as an AI coding assistant. Everything in the repo was created during the event.

Challenges we ran into

  • "Learning styles" is a myth. My first idea was to let students pick visual, auditory or reading "learning styles." Research doesn't support matching teaching to a preferred style, so I reframed the options as formats tied to real access needs (dyslexia, low vision, overwhelm, language) that anyone can switch between.
  • One visual doesn't fit every age. My first visual was a cartoon road with a mascot. It was great for a 7-year-old and patronizing for a high schooler, so I rebuilt it to depend on both age and assignment type.
  • Helping without doing the work. It would be easy for an AI to just write the essay. The prompt forbids it, the quick check tests understanding of the task, and essay evidence comes as ideas for kinds of sources, with a reminder to check every source before citing it.
  • Unpredictable AI output. If the model returned an icon or visual type the page couldn't draw, the plan broke. The server now pins every value to one the page can render.

Accomplishments that we're proud of

  • One assignment becomes a different, age-appropriate tool for each grade level, from a picture road for a preschooler to a synthesis matrix for a college student.
  • Share links carry the whole plan inside the URL, so a teacher can decode once and a whole class can use it for free, with no login.
  • It still works with no internet, no account and no money: saved plans, examples, read-aloud and printing all work offline.
  • Every format comes from a specific barrier, not a feature list.

What we learned

  • Accessibility features are most useful when each one is tied to a specific barrier, not added to a checklist.
  • Designing for "no internet, no account, no money" led to the best idea in the project: putting the plan inside a share link.
  • Telling the AI what not to do (no doing the work, no invented sources) mattered as much as telling it what to do.

What's next for Assignment Decoder

  • Test with real students and teachers.
  • Complete a full screen-reader pass.
  • Have native speakers check the translations.
  • Add a class view so a teacher can decode once and share it with a whole class from one page.

Built With

Share this project:

Updates

Submission history