A three-minute walkthrough.
What to click, in what order, and what each part is trying to prove. Written for anyone evaluating this — but it works just as well if you have simply landed here and want the short version.
Nothing here needs an account, a login, or an API key. Every page works on a phone.
The one idea, first
A tool that helps neurodivergent students specifically requires you to identify as one in order to use it. That means disclosing, and being marked out at exactly the moment you least want to be.
So Common Ground is built the other way round: a tool a whole class uses, whose design removes barriers that happen to fall hardest on some people.
Nobody has to disclose anything to benefit.
The walkthrough
- Step 1
Read a brief — 40 seconds
Go to Read a brief, press Use an example, then Read this brief.
The example is an ordinary-looking group assignment. It has 18 checks run against it and comes back with roughly twenty things worth asking about.
Look for: findings that match text quote the exact words that triggered them, so you can disagree with any one. Checks that look for something missing have nothing to quote, and simply name what is absent. Nothing is a black box, and nothing is a judgement about the reader — the summary says so explicitly: “None of this is a gap in you. It’s a gap in the brief.”
- Step 2
Carry it into the split — 40 seconds
In the results, find The parts this splits into and press Use these to split the work. Add two or three names, tap a couple of preferences each, and press Work out the split.
Look for: scroll past the roles to How this group works. If one person asked for several days’ notice before meetings, the group’s rule is now several days’ notice — with no name attached to it. Try it with different answers and watch the rule change.
That is the whole mechanism of this project in one screen: someone gets what they need without ever having to say why. - Step 3
Decisions — 30 seconds
Open Decisions. Add two names, pick which one you are, and record a decision.
Look for: it shows who has agreed and who has not answered, as its own state — and says plainly that those are not the same thing. Silence is never counted as a yes. Raising something in writing is a first-class entry, so nobody has to win a turn in a live conversation to get something on the record.
- Step 4
The display controls — 20 seconds
Press Display in the header, on any page. Change the text size, switch spacing to Roomy, try a different typeface and high contrast.
Look for: the controls are in the header on every page, in the same place — not hidden behind a label that makes you identify as needing them. Settings persist and are applied before first paint, so the page never flashes in the wrong state.
On the typeface: evidence for specialised “dyslexia fonts” is genuinely mixed, so the interface says “pick whichever is easier for you to read. There is no correct answer.” A preference, not a remedy. - Step 5
Optional — the parts that prove the claims
Where the AI is, and where it deliberately isn’t
Two layers. 18 deterministic checks are pure functions over the text — no model, no key. On top, an AI layer extracts what the brief asks for, a checkable definition of done, and the separable parts of the work.
The app works with no API key at all. Clone it, run it with no environment file, and the checks still work — that is tested, not just claimed. The AI improves the output; it is never load-bearing.
The role allocation contains no model call, on purpose. If a group disputes the split, “the AI decided” is not an acceptable answer. So it is a deterministic algorithm where every allocation quotes the preference that produced it.
The AI never rewrites anything. It extracts and asks. It never restates the brief in different words and never writes any part of the assignment — a rule taken directly from ASAN’s finding that generative AI “added new ideas that were not in the text we wrote.”
What it will never do
- Never infers or reports anything about a person. No account, no profile, no score. The only input is a document — there is nothing to infer from. A tool that detected neurodivergence and told a teacher would be outing people without consent.
- Never monitors, times, scores or gamifies. No streaks, no timers, no progress guilt. Masking research links performance pressure to real harm, so none of the usual engagement mechanics appear anywhere.
- Never teaches anyone to mask. It changes the environment, not the person.
One automated test enforces this directly: no finding may ever contain clinical or diagnostic language. If someone later adds a check that says “autistic” or “diagnosis” in its output, the build fails.
Why it looks this calm
There is no animation on this site. Nothing moves, nothing slides in, nothing autoplays, and the hero illustration is a static drawing rather than something that builds itself as you scroll.
That is a decision, not an omission. Sensory load is a documented barrier for a large part of the audience this was built for, and a tool that offers a “reduce motion” setting while animating its own landing page is not taking its own advice.
The effort went instead into things that cost nothing to read: generous line spacing by default, a short measure so lines never run too long, colour that is never the only way meaning is carried, and severity badges that always spell out the word.
Everything else
Source, the research it is built on, and the full reasoning behind each decision are in the GitHub repository.