-
-
Focus mode with large text: only the question and the answer stay on screen.
-
How do I cancel my place? The numbered answer card: the bottom line with its key facts marked, a tick-off checklist, the deadline.
-
How long until that link expires? Two links could match, so it asks which one instead of guessing.
-
The Chinese interface and a Chinese answer; the evidence keeps the English source, labelled English, not translated.
-
Who is the instructor? Not in the documents: the reader can ask the organisers to add it, with a consent box.
-
Reading settings in one drawer: focus mode, text size, bold word starts, read-aloud speed, colours.
-
The cover: a real rule is underlined, the one-sentence answer and its checklist slide out, and a Quote verified stamp is pressed in.
The problem
Practical student information (deadlines, eligibility, how to cancel, who to contact) usually sits in dense paragraphs of English. For people with ADHD or other attention, reading or executive-function difficulties, that format is a real barrier:
- Long paragraphs are hard to finish, so conditions buried mid-sentence get missed.
- Multi-step instructions are hard to hold in working memory.
- A deadline inside a paragraph is easy to overlook until it has passed ("time blindness").
- It is not obvious what the first step is.
AI summaries look like the fix, but they add a new risk: a fluent summary that invents or shifts a deadline does the most harm to the people who rely on it most.
Why it matters
Missing a registration deadline or a cancellation window can cost someone a place, a fee or an opportunity. The people most likely to miss these are often the ones dense text serves worst. Our rule: read less, but never be misled.
What it does
- A cover that shows the idea before you read a word. A short scene plays once on a paper desk: a real rule is underlined, the one-sentence answer and its checklist slide out, and a "Quote verified" stamp is pressed in. Example questions on the cover open the help desk with the answer already coming.
- The same numbered layout for every answer. A status banner says what happened (found / not in the documents / one question first / held back). Then ① the bottom line in one large sentence, ② what to do, ③ more detail (folded), ④ the evidence. Readers always know where to look. The key facts in the bottom line (date, time, place, contact) get a highlighter stroke and are repeated underneath as labelled chips, but only when the same values appear in a verified quote.
- A checklist you can tick. The steps the documents require, each starting with a verb, each with its deadline in a highlighted label that is never cut off, even on a phone. One click copies the list.
- Reading settings in one place: focus mode (hides everything except the question and the answer), three text sizes, optional bold word starts, reading-aloud speed, and light or dark colours. Choices are remembered.
- Listen. Reads the bottom line and the steps aloud in the browser's own voice.
- One language at a time. A 中文 / EN switch changes every label, hint and example question. Ask in Chinese and get the answer in Chinese; the only English left is the source quote, labelled "English, not translated", with the matching words highlighted as proof.
- Questions grouped by topic (dates, registering, who can join, help) instead of a wall of examples, plus a "try a tricky one" tab.
- Asks instead of guessing. If "that link" could mean two different links, it asks which one. If the documents don't cover something, it says so and gives the documented contact.
- Readers can ask for missing information to be added. A short consent-based report goes to an organiser dashboard, which shows what the documents are missing.
- It works without an AI key. A second engine, built the way help desks worked before large language models, turns the question into keywords (with a Chinese-to-English glossary), scores every sentence of the documents, and quotes the best one as the bottom line, with the steps and deadline lifted from the same sentence. It never paraphrases or invents. Without a key it answers by itself; with one, the reader can switch between "AI answer" and "Local engine".
- When the documents don’t answer, it drafts the email. The documented support address, the question as asked, and what the support guide asks a request to contain, ready to copy or open in a mail app. Nothing is sent.
How it keeps answers honest
- The model (Claude Haiku 5.5) only sees the club's current documents. An outdated draft is never sent to it.
- It must return structured JSON: a bottom line, steps, and an exact quote for every fact and every step.
- The server checks every quote word for word against the cited section. If nothing can be verified, the answer is withheld.
- Deadline check: every number in a step's deadline (day, month, year, time) must appear in that step's verified quote. This works across languages ("2026年11月12日 19:00" vs "12 November 2026 … 19:00"), and a changed date is flagged in red.
- Hidden instructions inside a question ("ignore the documents", "invent a phone number") are not followed.
How we built it
We reused open-source work wherever we could and wrote only the glue:
- Interface: React and Mantine (accessible components and charts), i18next for the two languages, Tabler Icons, built with Vite and TypeScript.
- Visual design: the "Coral" theme from the open-source Hallmark design skill (warm paper, letterpress controls, one coral ink for the verified state), motion following Emil Kowalski's design-engineering rules (short, purposeful, off under reduced motion), and Impeccable's quality checklist as a final pass.
- Key facts: chrono-node finds dates and times in the bottom line; rough-notation draws the highlighter stroke.
- Answer style: rules adapted from the MIT-licensed i-have-adhd project.
- Read-aloud: easy-speech (MIT), on top of the Web Speech API.
- Bold word starts: text-vide (MIT).
- Review queue: lowdb and csv-stringify.
- Evaluation: promptfoo.
- Accessibility testing: axe-core.
- Server: Node.js and Express, calling the model through the official Anthropic SDK with structured outputs.
- Design guidance: the W3C guidance Making Content Usable for People with Cognitive and Learning Disabilities (COGA), with WCAG 2.2 AA as the baseline.
What we measured (9 Nov 2026, fictional demo data)
- 40-question comparison. We compared our prompt with a plain "answer from the documents" prompt. Model, documents and server checks were the same; only the instructions differed.
- On 20 check questions, all automatic checks passed for 19/20 vs 15/20 in the first run and 20/20 vs 17/20 in a second run. The plain prompt's score moved between runs, so small gaps can be noise; the direction held.
- The difference was deadlines: the plain prompt attached deadlines that its quotes did not state.
- Reliability and cost. 214 real model calls by 9 November (smoke tests, questions typed into the app, and the evaluation), with 0 errors and 0 failed quote checks. All of them cost about US$0.13 in total, well under 0.1 US cents per question.
- Tests. 35 automated tests pass, including the local engine and the email draft.
- Local engine, 34 demo questions (English and Chinese): the right sentence for 29, a clarifying question for 2, "not covered" for 3, nothing invented.
- Accessibility. axe-core (WCAG 2.2 A/AA plus best-practice rules) found 0 violations on every screen, including the cover, in Chinese and English, light and dark, on desktop and at phone width (375 px).
These are automatic checks on a small fictional dataset, not an accuracy figure.
Hand-scored review (11 Nov 2026). A person scored all 40 answers from each prompt against the documents: the facts were right in 40/40 TraceDesk answers and 34/40 plain answers, and TraceDesk’s bottom line was faithful to the documents in 40/40. Three of the plain prompt’s six wrong answers had passed every automatic check.
Challenges
- Shorter but faithful. Answers had to get shorter without drifting from the source. Our fix: every checklist step needs its own verified quote, plus the cross-language deadline check.
- A bug our evaluation found. Our deadline check first compared digits only, so a correct Chinese date was flagged red against "12 November". We fixed it and added regression tests.
What's next
- Test with volunteers, with consent, anonymously and without collecting health data, and change the design based on what they struggle with.
Honest notes
- The documents belong to a fictional club ("DemoClub"), not a real service.
- It has not yet been tested with people with ADHD. We say it was "designed with" cognitive-accessibility guidance, not that it is "proven to help". It is not a medical tool and records no health information.
- All application code was written during the hackathon, from 9 November 2026. The only material that existed before the event (3 November) is the fictional documents and the draft test questions.
- Built by our team with help from an AI coding assistant (Claude Code). Reused components are credited in THIRD_PARTY_NOTICES.md.
Setup
Requires Node.js 22+ and an Anthropic API key.
- Run
git clone https://github.com/Jason-d-e/tracedesk-lite, thencd tracedesk-lite/app. - Run
npm install. - Copy
.env.exampleto.envand put your key afterANTHROPIC_API_KEY=. - Run
npm start, then open http://localhost:3000 (help desk: /#/ask, organiser dashboard: /#/organiser).
Built With
- anthropic
- axe-core
- claude
- express.js
- i18next
- javascript
- lowdb
- mantine
- node.js
- promptfoo
- react
- typescript
- vite
- web-speech-api