How our AI works

The one rule

Writ’s AI never writes any of your personal statement. Not a sentence, not an opening line, not an “example” to adapt. It asks the questions an admissions tutor would ask, points at what is weak and why, and leaves every word to you. Your statement has to be your own work under UCAS’s rules, and because none of the words come from us, nothing of ours can turn up in it.

Which AI it is

Writ uses Claude, made by Anthropic, currently Claude Sonnet 5.5. If it is busy, or declines a request, the same request goes once to Claude Sonnet 5 so your round still comes back. Anthropic processes what we send it to produce the reply and, under our agreement, may not use it to train its models.

What it does

• The coach reads one section of your statement and asks up to three questions, only as many as there are real gaps. It names what is already strong so you keep it, flags what a sceptical reader would press on, and gives you a list of ingredients for the section (never sentences). When more rounds would not help, it says so.
• The whole-statement review reads all three sections together against what your subject looks for. If you are over UCAS’s 4,000 characters, it points at a few cuts in your own words; it never makes them for you.
• The profile analysis reads what you have done and recommends super-curriculars, only from our checked catalogue, with advice on how to fit them in. It is checked so that no line of it is written in your voice.

What it reads

To coach one section, the AI is given: that section, your other sections, your answers to its earlier questions, your subject, predicted grades and five choices (with what those courses teach), what you told us in onboarding, and any notes from your tutor.

It is not given your email address, your payment details, or your note about what your grades don’t show. That note is read only by the profile analysis, to advise on contextual offers, and never reaches the statement coach.

How we hold it to the rule

• The instructions it follows forbid writing statement text, and say why: a borrowed sentence would not survive a question about it.
• What comes back is checked by code before you see it. Answers in the wrong shape are refused, and advice phrased as a line you could paste (“I first became interested in…”) is removed.
• Before every change of model we test it with students’ tricks: pleading about a deadline, “my teacher said it’s allowed”, instructions hidden inside a draft, asking for “just one example”. It has to refuse them all.
• A human tutor reviews your finished statement if your plan includes it, and their notes outrank the AI’s.

Where it can be wrong

It is a tool, not an admissions tutor. It can misread a draft, ask a question that misses the point, or be more confident than it should be. It does not know whether you will get an offer, and nothing it says is a promise. Course figures come from official sources (where our data comes from), but always confirm entry requirements on the university’s own page. If something it says seems wrong, it probably is worth questioning, and we would like to hear about it: enquiries@writai.co.uk.

Does it learn from my statement?

Not unless you say yes. Using your statement to improve Writ, including ever training an AI of our own, is a separate permission that is off by default. See your permissions.