This is a submission for the Presage SmartSpectra SDK Challenge.

What I Built

Every writer who has sent a chapter to beta readers knows the reply: "Loved it!" It is kind, and it tells you nothing about the paragraph where they stopped reading in their head and kept turning pages. Unskipped finds that paragraph. A beta reader who agrees to it reads your chapter on their laptop with the camera on, and the Presage SmartSpectra SDK reads their pulse, blinks and smiles on their own machine. Every note they choose to share lands as a Word comment on its paragraph, signed with their name, in your own manuscript. The webcam pulse it builds on held up: against a fingertip pulse oximeter, 29 of 29 readings were within 5 bpm.

Elmore Leonard's tenth rule of writing is "Try to leave out the part that readers tend to skip" (The New York Times, July 16, 2001). I built this for a novelist revising chapter by chapter, who sends each chapter to three beta readers and gets back three versions of "loved it". The reader in this post is me, reading a story I had never read.

Unskipped turns what the camera sees into one mark per paragraph, like the translucent page flags a reader sticks in a book:

  • pink: gripped (their pulse rose, or their blinking slowed down),
  • yellow: they smiled,
  • blue: they may have drifted (their blinking sped up: in a 2010 study at the University of Waterloo, readers blinked more when their minds wandered from the page),
  • clear: no reading. They looked away, talked, or the light was bad, and Unskipped says so instead of guessing.

The reader agrees first, on a screen that says in plain words what is measured and that nothing leaves their laptop until they press Send:

The consent screen

They read one paragraph at a time, so the app always knows which paragraph their face and pulse belong to. Paragraphs ahead stay faded until they reach them:

Reading, with the live SmartSpectra readings beside the page

Before they see a single flag, they mark the paragraphs that got them and lost them themselves. Then they review every note, drop any they don't want to share, and send a small file. The writer adds it to their manuscript, and each note becomes a Word comment on its paragraph. It opens in Word or Google Docs, where the writer already works.

The review screen: the chapter seen edge on, flags, and the note beside its paragraph

The exported manuscript in Word: each kept note is a comment on its paragraph, signed with the reader's name

Demo

Try it with no camera and no key: the replay page plays back my real recorded reading through the same scoring code: every flag, every note, the numbers. From there you can also walk through every screen with simulated signals.

Code

GitHub logo JonathanSolvesProblems / unskipped

Your beta reader's face and pulse leave page flags in your manuscript. Presage SmartSpectra on the reader's laptop, notes as Word comments. DEV Presage challenge entry.

Unskipped

Your beta reader's face and pulse leave page flags in the margin of your manuscript.

A reader who agrees to it reads your chapter on their laptop with the camera on. The Presage SmartSpectra SDK reads their pulse, blinks and smiles, and whether they are talking or looking away, on that laptop. Unskipped turns it into one flag per paragraph:

  • pink: gripped (pulse up, or blinking slowed),
  • yellow: smiled,
  • blue: may have drifted (blinking sped up),
  • clear: no reading (looked away, talking, poor light). It says so instead of guessing.

The reader marks the paragraphs that got them before seeing any flag, reviews every note, keeps or drops each one, and sends a small file. The writer adds it to their manuscript, and each note becomes a Word comment on its paragraph, signed with the reader's name.

Entry for the DEV Presage SmartSpectra SDK Challenge, November 2026.

  • Demo…

Setup is in the README: npm install, put a free Presage key in .env, npm start. npm test runs the unit tests, npm run e2e drives every screen in a real browser, and npm run check re-scores the committed reading sessions and fails if any number in this post disagrees with them.

How It Responds to Human Signals

SmartSpectra signal What Unskipped does with it
Pulse rate (with its confidence) Each confident reading is credited to the paragraph read about 6 seconds earlier, since pulse is an average over roughly 12 seconds. A paragraph well above the reader's own median gets a pink flag.
Blinking Blink rate per paragraph against the reader's own rate, judged only where that rate predicts at least 3 blinks. Well under it: absorbed (pink). Well over it: may have drifted (blue).
Expressions A happy expression held for 0.7 s or more: yellow.
Talking Time spent talking is not reading time, so a paragraph read mostly aloud gets "no reading".
Validation status, face mesh No face, face not forward, too dark, moving, or a head turned well away from the reader's usual position: no guess for that paragraph.

Everything is measured against the reader's own baseline for the session, never against an average person. Guided reading came out of testing: my first version let the reader scroll freely and credited whatever paragraph sat under a reading line, and in a pilot read a 9-word paragraph got 266 seconds because I read a whole screen at once.

How well it works

The pulse. I wore a fingertip pulse oximeter (LOOKEE A310) for 2.5 minutes while the laptop webcam measured my pulse, and filmed its display with my phone. Lined up every 5 seconds: the camera averaged 77.1 bpm and the oximeter 77.3, a mean gap of 1 bpm, with 29 of 29 pairs within 5 bpm.

The webcam pulse against a fingertip oximeter

Against my own marks. I read Saki's "The Open Window" (1914, 29 paragraphs), then marked the paragraphs that got me before seeing any flag. The camera's flags found 2 of the 4 I marked (flags placed at random would find 1.5), and time spent on the page found 3 (by chance: 1.1). The review screen shows both, because with one reader neither is enough alone.

Limits, and what's next. Resting pulse wanders by about 3.5 bpm, so the current pulse bar (4 bpm) also flags a reader with no reaction at all on up to 22.3% of paragraphs (node tools/null-check.js, using my real resting pulse). Raising it to 8 bpm cuts that to 6.5%; that is the next change, to be tested on new reads rather than fitted to mine. In planted controls, written down before the run (look at my phone, read aloud, turn away, and two paragraphs read normally), my first session caught 1 of 3 for the right reason: the SDK kept seeing a face while I turned. With head turn added from the SDK's face mesh, a second session caught 2 of 3 and read 2 of 2 normal paragraphs. The miss was a glance down at my phone with my eyes alone: the SDK read the lowered eyelids as blinking, so that paragraph came out "may have drifted" instead of "no reading". Every threshold, every change and every session is in the repo (results/controls.md, results/README.md).

Prize Categories

Best Thoughtful Real-World Use. The category asks for "SmartSpectra applied to a clear real-world need with strong usability, consent, and responsible data handling."

  • Real-world need: every beta read of every chapter, for every writer who has heard "loved it".
  • Usability: one paragraph at a time, marks before results, and the output lands in the writer's own Word file.
  • Consent: a screen in plain words before the camera turns on, a reader who approves every note before anything is sent, and Stop on every screen with the camera on.
  • Responsible data handling: processed on the reader's laptop, telemetry off, no server, nothing kept after the reading is sent (below).

Most Creative or Delightful Use. A reader's pulse, blinks and smiles leaving page flags in a manuscript, read back as Word comments.

Consent and Data Handling

  • Consent. The camera does not turn on until the reader ticks "I agree to be measured while I read this piece" on a screen that lists, in plain words, what is read, where it is processed, what the writer will get, and that it says nothing about their health. Nothing is measured until they press Start reading. Stop is on every screen with the camera on, and closing the window turns the camera off.
  • Processing. SmartSpectra runs on the reader's computer (the Node.js SDK in the app's main process). No video or image is recorded, saved or sent by Unskipped.
  • What goes over the network. The SDK contacts Presage to authenticate the API key and to count usage during a measurement. That is all the network traffic Unskipped causes.
  • SDK telemetry. Off. Unskipped starts the SDK with enableTelemetry: false; Presage's docs say that with it off, "the SDK starts no telemetry session and transmits nothing" for telemetry.
  • Storage. During a reading, the signal samples (pulse value and confidence, blink and talking flags, a smile score, whether a face was found, and two head-position ratios; no image and no face geometry) are kept in memory and in a crash-safety backup file in the app folder. The app deletes that backup as soon as the reading is sent or stopped. The file the reader sends contains only the notes they kept, their own marks and summary numbers; raw samples are included only if the reader ticks a box that is off by default.
  • Sharing. Nothing is uploaded anywhere. The reader saves a file and sends it to the writer however they like. The writer's manuscript never leaves the writer's computer.
  • Retention and deletion. Once a reading is sent, the only copies are ordinary files on the reader's and writer's disks; deleting them deletes the data. Unskipped has no server. Presage's authentication and usage records are governed by Presage's own policy.
  • AI services. None. No language model or other AI service is used by the app.
  • Not a health tool. The readings describe how someone read one piece. The app makes no health claim and uses no health language.

What's next

On 20 November I wrote to the Quebec Writers' Federation in Montreal and offered a free session to any member who wants to try it on a chapter: I set it up and sit in, their beta reader reads, they get the comments. I also asked what a writer would need to see before trusting something like this, because that answer decides what gets built next. After that: the higher pulse bar tested on new readers, more readers on the same chapter (the export already stacks several readers' comments on one manuscript), and a longer read than a short story, where drifting matters most.

Agent Session

I built this with an AI coding agent (Claude Code). The session is in the repo, readable on GitHub: agent-session. It is three stretches of the build, cleaned of addresses, keys and file paths:

  • Camera tests, 19 November. The laptop webcam gave no pulse at all. The SDK's own log showed why: "chest not visible" kept breaking the picture, and pulse needs about 12 seconds unbroken. The fix was framing (face and upper chest) and asking for PULSE_RATE alone, now in the README's SDK notes.
  • The Word export, 19 November. Checking that the exported comments open in Google Docs under the reader's name, on the right paragraphs.
  • A code review that changed the scoring, 20 November. Each pulse value placed at its own timestamp, the chance level added beside every grade, and a no-reaction simulation run before and after the fix.