Teleprompter: the prompter
follows the signer.

Every prompter ever built scrolls to a voice. This one finds its place by watching a signer's hands — the fingerspelled names, the numbers, a handful of distinctive signs — and snaps the script under the reading line as those land. Then it hands the very same script to a hearing co-host in voice. One script, two senses, no second tool to learn.

What Teleprompter is

A teleprompter is a scroll loop: a script crawls up a sheet of one-way-mirror glass tilted in front of the lens, the presenter reads, and the words keep pace with delivery. For a hundred years that pace has meant one thing — a speaking voice, or a foot pedal, or a dumb timer you fight the whole way down. None of them were built for a person who presents with their hands. Teleprompter is the one that is. We wrote about why nobody had built it in A prompter for signers; this is the build proof underneath that essay.

The hard part is figuring out, in real time, which word the presenter is on. A speaking prompter cheats: it transcribes the voice and keeps the scroll roughly under the next line. We do the harder version, and we do it without transcribing anything. The presenter is reading a script we already have, so the job is not to understand them — it is to find their place. The page watches for the words that are easiest to catch on a pair of hands — fingerspelled names, places, numbers, and a small set of distinctive signs — and snaps to them. A handful of those anchors per paragraph is enough to keep the script under the reading line. Whatever the page can't catch in time, the baseline pace simply carries; a missed word never freezes the scroll.

The second thing it is: universal by construction. The piece that reads a voice and the piece that reads a pair of hands are the same engine, reading the same script. A hearing co-host can run voice mode off the very page the signer just presented from — same words, same scroll loop, in front of the same room. We did not build a signing prompter and bolt voice on afterward, the way access usually gets added to a finished product. We built one prompter that doesn't care whether the next word arrived through a camera or a microphone, and the voice version is what falls out for free.

What we built

The split that makes the whole thing work: one alignment engine in the middle, two senses feeding it the same shape of signal. The engine never asks which sense a word came from — that is the architectural commitment that keeps it honest as universal design rather than a niche tool with a hearing fallback.

• Anchor alignment No full transcription — the page finds its place by spotting the words easiest to catch: fingerspelled names, places, numbers, and a small set of distinctive signs. When one lands, the script snaps to it; a sliding window keeps the match honest and refuses to jump backward on a stray false hit. Sparse paragraphs ride the baseline pace until the next anchor lands.
• Hands-free start and stop Starting and stopping has to be hands-free too, or the presenter is reaching for a keyboard mid-take. Hold I-L-Y toward the camera for three seconds and the session counts down and runs; hold it again and the session ends. The three-second hold is the whole safety: it is a sign no presenter holds for three full seconds by accident, and a progress ring around the marker shows the hold registering.
• Mirror mode + eyeline marker Built for camera-direct delivery. A mirror mode flips the text for presenters who shoot through prompter glass, so the reflection reads correctly to the eye. An eyeline marker sits at the upper third, holding the reading line where the lens wants the gaze — so the presenter looks at the camera and reads at the same time, the way the old glass trick always intended.
• Voice mode, same script The hearing co-host's path. It consumes a streaming caption service and feeds the exact same alignment engine the words it hears — off the exact same script the signer used. Nothing about the prompter loop changes between the two; only the sense feeding it does. The feature usually metered hardest is the one that stays unlimited here for the signer.
• A rest is a section break A signer doesn't speak commas — they mark structure with their body, and the clearest body-level signal is stopping between thoughts. After a stretch of steady signing, a brief pause snaps the page forward to the next paragraph. It is the highest-confidence non-word signal the hands produce, and it costs the presenter nothing they weren't already doing.
• The camera never leaves the room Sign recognition runs entirely on the page in front of the presenter — the camera frames never go anywhere. Every recording state carries color, an icon, and text together, captions are on by default, and the whole thing still works as a plain manual-scroll prompter for anyone who declines the camera and microphone. That floor is not a setting; it's the design.

* Design target. The three-second hold is the false-positive guard for the start/stop sign; the anchor spacing is the working assumption the engine is tuned around — roughly one anchor every thirty words keeps the script snapped to position, and the baseline pace covers the gaps. Both are tunable per script, not guarantees of any one session.

Where the engine came from

This is the part worth being honest about. We did not invent hand-tracking on a whiteboard. The hand-reading the prompter leans on is the same recognizer that already drives 1891's signing demos — the part that watches the camera, on the page, and reports the letters and signs it is confident about. The good idea was noticing that a prompter doesn't need to understand ASL to follow a signer. It needs to find a place in a script it already holds. That is a far smaller problem than recognition in general, and it is exactly the problem the recognizer is already good enough to solve.

So the engine sits one layer above the senses. A normalized event — a word, a confidence, a timestamp — arrives, and the engine doesn't care whether a camera or a microphone produced it. The sign path wraps the on-page recognizer; the voice path wraps a streaming caption service behind a thin proxy so the presenter's page never holds a vendor key. Same engine, same script, either way. Branching the alignment logic on which sense fed it would have been the mistake; we wrote it so we never could.

The prompter follows the signer. Not the other way around. — a working principle, written into the engine on day one

Read a script.

Below is a small, working slice of what the alignment engine does. Press Advance to deliver the next anchor word — the kind a pair of hands catches easiest — and watch the page snap that word to the reading line, then resume its baseline crawl. Between anchors the script keeps scrolling at the set pace, exactly the way it does when the presenter is mid-sentence and nothing has landed yet. Flip on Mirror to see the glass version, or switch to Voice to feed the same engine spoken words off the same script. Nothing here phones home; it's the snapping logic, running in your browser.

Why this fits 1891

The pitch on madeby1891.com is for systems built end to end, in Frederick, under one roof, and built so the person at the front of the room can run them. Teleprompter is that pitch pointed at a podium. The presenter keeps their hands — no clicker palmed out of frame, no pedal to find blind, no assistant in the wings guessing when to nudge the scroll. The voice mode isn't a lesser fallback bolted on at the end; it is the same engine, off the same page, so a Deaf presenter and a hearing co-host share one tool instead of two. None of that is an accessibility skin. A prompter that starts from the signer at the podium is just what the room needs to work for everyone in it.

There's a longer line behind it, too. 1891 is a family trade as much as a product line — five generations Deaf, the ones before standing at the front of mixed rooms long before there was a tool that assumed they'd be there. A prompter built so a signer never loses their place is a small thing measured against that. It is also exactly the thing nobody bothered to make. The cameras are already in the room and the script is already written; what was missing was a prompter that begins with the hands and lets the voice ride along, instead of the reverse.

What's still unfair about this

We are honest about the edge. The prompter follows by anchors, not by reading every word, so a paragraph of common verbs and pronouns gives it little to grab — there the baseline pace carries the scroll and the presenter is, briefly, reading a plain prompter. That is by design: a missed anchor must never freeze the page, and a stray false hit must never yank it backward through the script. Best-effort alignment that never breaks beats perfect alignment that sometimes stalls a presenter mid-take. We would rather under-promise the catch and never lose the scroll.

The thing we did not do is frame this as ASL recognition. It isn't. The page spots a small set of distinctive signs and fingerspelling as anchors; it does not understand grammar, classifiers, or the face. Calling that "ASL recognition" would be overclaiming, and the copy never does — it follows along, it does not translate. The presenter brings their own language; the prompter just keeps their place. We measure it by whether the signer kept their hands and never lost the line, not by how much of the language we could claim to have read.

Nobody at the podium cares about the engine. They care that the script was under the line when they looked up, that they never reached for a button, and that the co-host ran the same page without a second tool. — Anthony Mowl, framing the build for Teleprompter

That sentence is the spec. The prompter is good because it disappears.