# Cue Scroll

> A rehearsal teleprompter for people who speak in public: rehearse a speech at the pace you mean, see where you rush or drag, and have a model mark it up for delivery and trim it to the time you have. https://cue-scroll.skillsafe.ai/

Cue Scroll is for a wedding toast, a eulogy, a keynote or conference talk, a sermon, a lesson opener, a team presentation or a pitch you have already written. You bring the script (typed, pasted, or dropped as .txt, .md or a Cue Scroll .json) and the time you have. The work object is the delivery, not the writing: the app never rewrites your sentences.

## What runs free in the browser

- Paragraphs and sections: a blank line starts a paragraph (ids p1, p2, ...); a line starting with ## starts a section.
- A mark grammar you can type or insert from the toolbar: `/` pause (0.6 s), `//` long pause (1.6 s), `^` breath (0.4 s), `*emphasis*`, `[slow]` (words to the end of the sentence take about 18% longer), `[cue: text]` for a slide or stage cue.
- Timing: planned seconds per paragraph and per section at your words per minute (pauses included), fitted seconds scaled to your target, and the pace you would need to fit the target as written.
- For the ear: sentences over N words (default 25), Flesch reading ease per paragraph for English, sentences stacking three or more numbers.
- The teleprompter: near-black stage, very large type, a fixed reading line, smooth scrolling at exactly your words per minute (each rendered line moves at the time span of its words), mirror mode for a glass prompter, font size and margin controls, full screen, and the keyboard: Space play/pause, Up/Down pace, Left/Right paragraph, M mirror, F full screen, Esc stop.
- Rehearsal mode: the stage scrolls at the planned pace and holds at each paragraph end; you press Enter (or the big button on a phone) as you finish each paragraph. Double taps under 350 ms are ignored, paused time is excluded, Backspace undoes a tap. Each paragraph is then compared with its plan: under 88% of the planned time is rushed, over 112% is dragged (with a one-second tolerance for short paragraphs). The last eight rehearsals are kept per script.
- Exports: the marked script as Markdown (pending cuts shown as strike-through), printable cue cards (one paragraph per card, two per page), paragraph timings as CSV, and everything as JSON.

## The model lanes (metered, gpt-terra)

- `mark`: delivery marks for every paragraph, trims to the target (each an exact contiguous run of words from one paragraph id, with the words saved), the arrow (the one sentence the audience must remember), how to open and how to land the close.
- `remark`: after a rehearsal, re-marks only the paragraphs still rushed or dragged, with one concrete pacing instruction per paragraph.
- Every reply is reconciled before anything lands: paragraph ids must exist; marked text must keep the speaker's wording (small drift: the marks are moved onto the original words; a rewrite: the marks are refused); a cut must appear word for word; words and seconds saved are recounted at your pace; the plan with the coach's marks and after every cut is compared with the target. Cuts are strike-through ghosts in the prompter and in the paragraph list until you accept them one by one.

## Input contract

One JSON object, scalar fields: `task` (mark | remark), `title`, `occasion`, `target_seconds`, `wpm`, `slides`, `planned_seconds`, `paragraphs` (JSON-encoded array of {id, words, secs, full, text}), `rehearsal` (JSON-encoded array of {id, planned_secs, actual_secs, pace} or ""), `targets` (comma-separated ids, remark lane), `clipped`. Long scripts are clipped on paragraph boundaries: the first two, the last two, the targets and the rushed or dragged paragraphs stay full; the rest are sent as their first 30 words with a count, and are never marked or cut.

## Developer surface

- API tutorial: https://cue-scroll.skillsafe.ai/api.html (task field, input fields, reply contract, worked examples, eight languages)
- The timing engine, recorder and reconciler are a plain script: https://cue-scroll.skillsafe.ai/cuekit.js

## Source

Built on the SkillSafe skill public-speaking (@refoundai/public-speaking, from Lenny's Podcast guests): decompose the skill, embrace the pause, the Bow and Arrow method (one sentence the audience remembers), End Strong, and the Accordion method of compressing a talk to its time.