Plain English — What Does
This Letter Actually Say?

// the idea

Everyone gets letters they can't quite read

A note from the hospital with a code in it. A clause in a lease. An insurance decision that uses the word "adjudicated." A "we've updated our terms" email that is eleven paragraphs long and says one thing that matters. These are written by people who do this all day, for people who don't, and the gap is not an accident — it's just that nobody on the sending side is paid to close it.

This is the most obviously useful thing a language model can do for an ordinary person, and it's underserved because it's not glamorous. Plain English is a translator. Paste the confusing text, say roughly what it is, and get back what it actually says, what it means for you, the questions worth asking whoever sent it, anything that deserves a second look, and the right kind of person to talk to next. It does not tell you what to do. It makes sure you understand what you were told, so that the deciding is yours.

A concrete example

The built-in example is a real-shaped insurance letter — the kind that arrives after a scan:

…a determination has been made that the services billed under CPT 73721
are not eligible for reimbursement under your plan's Schedule of Benefits,
as prior authorization was not obtained in accordance with Section 4.2 of
your Evidence of Coverage. As such, the allowed amount has been applied as
a member responsibility. You may be entitled to request a redetermination…
Requests for a first-level internal appeal must be received in writing
within 180 calendar days of the date of this notice…

And what comes back:

Summary  Your insurer has refused to pay for the knee MRI because nobody
         got their approval in advance, which your plan requires. They're
         saying the cost is now yours — but you can appeal, and you have
         180 days to do it.

Ask      Who was responsible for getting prior authorisation — me or the
         clinic?  ·  Can the clinic submit a retroactive request?  ·  What
         is the "allowed amount" in dollars?

Flags    The 180-day clock started on the date on the letter, not when you
         read it.  ·  "Not a bill" means a separate bill is coming.  ·
         Appeals must be in writing — a phone call won't count.

Jargon   adjudicated — they've processed and decided on it
         CPT 73721 — the billing code for a knee MRI
         member responsibility — the part you're being told to pay

Nothing in that is advice. Every line is either a plain restatement of something in the letter, a question the letter leaves open, or a pointer to who can answer it. That's the discipline the whole tool is built around.

// what it does

Two panels

// how it was built

Build notes

The server piece is deliberately tiny

Unlike the other tools on this site, this one can't be key-free: nobody pasting a hospital letter should have to own an API key. So there is a PHP proxy, the same shape as On This Day's — about ninety lines. It reads the text, kind and situation, sends one request to Claude, and returns the result. The key lives in an environment variable or a file outside the web root. There is a per-IP throttle of six requests a minute using a temp file keyed on a hashed address, so one visitor can't run up the bill. Nothing is logged.

The prompt is a job description, not a persona

The system prompt tells the model it is a translator, not an adviser, and gives it rules rather than a character: use only what's in the text, say when something's missing instead of guessing, quote the actual number or date, never invent a law or policy, never claim to be a lawyer or doctor, and flag anything that looks like a scam. It asks for a fixed JSON shape so the page can lay the answer out properly; if the model wanders and returns prose, the proxy passes it through and the page shows it raw rather than failing.

You translate confusing official text into plain English for an ordinary
person who has to deal with it. You are a translator, not an advisor.

- Use only what is in the text. If something important is missing, say so.
- Be concrete about deadlines, amounts and rights the text states.
- Do not tell the person what to decide. Do tell them what it means for
  them, what to ask, what looks unusual, and who to talk to next.
- Never invent a law, policy or fact not in the text.
- If the text looks like a scam or a fake demand, say so clearly.

The scrubber runs before the send, in the browser

The personal-details button is a dozen regular expressions — emails, phone numbers, SSN shapes, long digit runs, "account/claim/policy number" followed by an identifier, street addresses, salutation-plus-name, dates. It is not perfect and doesn't claim to be, but it runs entirely in the page and it removes the things that would make a pasted letter identifiable. The privacy note says exactly where the text goes and why the button matters.

// what this demonstrates

Why this matters

Most of what's written about AI is for people who build with it. The case that actually persuades everyone else is a tool that does one ordinary thing well, says clearly what it won't do, and treats the person's information with care. A parent who gets a confusing letter and finally understands it is a better argument for the technology than any benchmark.

It's also the counterpart to Promptwatch: that one is about keeping a prompt honest as it changes, this one is a prompt that has to be honest from the first version, because the person reading the answer can't check it. The rules in the system prompt above are the regression tests, written in English.

Where it could go next