data-reels

Files (1)

• SKILL.md

SKILL.md (first lines)

# Data Reels

Principles for turning quantitative content into short-form video that holds attention, distilled from measured competitor research (view-to-follower ratios, outlier analysis, transcripts) across data-viz and stats Instagram accounts. This skill is the philosophy layer — the "why" and the invariants. Your project's own reel playbook (e.g. `digital_marketing/REEL_PLAYBOOK.md`) is the operational layer — the "how": render pipelines, asset-fetch scripts, exact CLI commands. Reference that project playbook for execution; use this skill for judgment calls.

## 1. The rejection-cascade / one-insight format

The best-measured format is not a countdown. It's:

- **0-2s — thesis claim.** State the plain-language claim or question immediately. No throat-clearing, no "hey guys," no title card that isn't the claim itself.
- **3-4 beats — assumption → correction, with visual proof each time.** Each beat states a plausible-but-wrong read, then shows the visual that rejects it ("X? No — [proof]"). Screen recording, animated chart, or overlay — never just narration.
- **Close — one crisp rule, no CTA.** A single quotable takeaway that could stand alone as a tweet. No "follow for more," no cliffhanger, no ask.

This beat structure measurably outperformed 5-item countdown formats in the same niche. The whole video *is* the payoff, delivered as a sequence of corrections — not a tease building to a reveal.

## 1.5 Format follows story shape

The rejection-cascade (section 1) is not the only valid format — it's the right format for one
specific story shape (a contradiction or anomaly being corrected in a sequence of beats). Match
the format to the story, don't force every insight through the same template:

- **Contradiction/anomaly** ("this looks like X, it's not") -> rejection-cascade: scene-swapped
  beats, each stating a plausible-but-wrong read then rejecting it with proof.
- **Comparison/duel** (A vs B, "bigger than," "beats") -> **one composition evolving, not
  scene-swaps.** Build a persistent split/scale layout that stays on screen the whole reel;
  elements transform in place — a number ticks up, a bar or beam tips, one side visibly
  overtakes the other — rather than cutting between two separate scenes for A and B. The reel's
  entire job is showing one continuous transformation, not a sequence of discrete reveals.
- **Ranked list** (top N) -> countdown, frontloaded: put the strongest item first, not last —
  short-form retention drops before a slow build to a big finish pays off.

The generalizable principle: a scene-swap format asks the viewer to remember beat 1 while
watching beat 3. A comparison story doesn't need that memory load — both sides can just stay
visible and change value/position in front of the viewer, which is both easier to follow and
better matches what a comparison *is* (a single relationship, not a sequence of separate facts).
When a story is fundamentally "A vs B, and here's the moment B overtakes A," resist the urge to
default to whatever format was used last — the composition-evolves shape will read as smoother
and better-paced than forcing it into cascading scene cuts, which read as rushed/rough for this
specific story shape (confirmed via user feedback on a rejection-cascade cut of a comparison
story that was rebuilt as a persistent-composition format afterward).

Pacing implication: a comparison/duel format needs room for a single climax to breathe (place it
around 60-70% through the runtime, give it the longest beat window) and strictly soft transitions
between beats (500-700ms crossfades/eases, never a hard cut) — since nothing ever leaves the
screen, an abrupt transition on any single element reads as jarring in a way it wouldn't in a
scene-swap format where cuts are the expected rhythm.

## 1.6 Format classes beyond rejection-cascade and head-to-head

Measured competitor research on out-of-niche data-story accounts (geography/history/map creators) surfaced additional format classes worth matching to story shape, beyond the two above:

- **Single-transformation reel.** One continuous, wordless, self-explanatory visual change (a map redraw, a silhouette shrink, a region wipe), 6-15s, with **no pacing scaffolding** — no chapter cards, no VO, sometimes not even a caption beyond a one-line title. The strongest measured example ran 6 seconds and hit 41x its account's own median view-to-follower ratio. **Key insight: duration is not the retention lever here — a single continuously-evolving self-explanatory visual is.** Don't assume every short-form idea needs beats, hooks, or narration scaffolding; if the transformation itself is legible and dramatic, let it carry the entire reel unassisted.
- **Timeline biography.** An era-scrub playhead traveling across a horizontal/vertical timeline with milestone markers popping in at key dates — the "life story of a subject" shape (a market, a product line, an institution). The persistent progress cue doubles as a completion-satisfaction driver.
- **True-size / nesting comparison.** A layered-silhouette overlay: a full-scale shape with a smaller shape shrinking/nesting inside it to true relative scale. Fits any "X is bigger/rarer/smaller than you think" claim — population counts, print runs, market caps as overlapping silhouettes instead of plain bar charts.
- **Bar-chart race.** Ranked bars swap order and animate width/position continuously over a long time span. **Duration caps do not apply when the animation itself IS the story** — a measured 175-second bar-race reel (vs. a typical 15-30s target) still hit a strong outlier ratio, because viewers watch a long unbroken race for the rank-churn payoff itself, not despite its length. This format only works when the racing entities are already familiar to the viewer (well-known cities, countries, brands) — it's a poor fit for a single unfamiliar entity with no "cast" to recognize.
- **Chapter story.** Self-aware, comedic "oversimplified version" framing with numbered chapter cards and a "part 2" tease at the close — lowers perceived stakes so viewers forgive compressed nuance. Carries genuine off-brand/tonal risk: it borrows a parody-documentary voice that may clash with a data-authority brand identity. Use deliberately, don't default to it.

## 1.7 Share/save triggers

Distinct psychological drivers behind why a data-story reel gets forwarded vs. bookmarked — worth naming separately because they call for different production choices:

- **Identity signaling.** Rivalry/team framing ("team X vs team Y," fandom-vs-fandom) lets a viewer self-insert and tag someone on "the other side." **Identity-signaling rivalries are the strongest share trigger for fandom data** — it converts a data claim into a social move (tagging a friend), which is a stronger distribution mechanism than the data being merely interesting.
- **Nostalgia arcs.** "Then vs now" / vintage framing triggers a documented dopamine/belonging response and reliably outperforms polished, hyper-modern content on feeds.
- **Completion satisfaction.** Timelines, countdowns, and races resolve an open information loop; visible progress (playheads, milestone markers) intensifies the pull to watch to the end.
- **Reference-utility end cards ("save this").** A reel that holds its final numbers static for a beat reads as a re-checkable artifact worth bookmarking rather than a one-time watch — long runtime combined with a low like-to-view ratio is the empirical signature of this behavior (people watch fully to save, not to react).
- **The "whoa" transformation gasp.** A single wordless visual payoff is forwarded on its own visual strength with zero domain context required — this is a share trigger, not a save trigger, and correlates with low like-to-view ratios (high reach via algorithmic pass-along, not deep engagement).

## 2. Consistency over virality

The best pure-data account measured had **zero statistical outliers** — its top reel was only 1.35x its own median. Every post performs; none spike, none flop. That is the deliberate trade: template consistency as the strategy, not "swing for a viral outlier."

Contrast: an account chasing viral highlight moments had one reel at 96x its own median — and that reel had nothing to do with the account's core format (it was a lucky highlight clip with a caption, not the data format at all). Outlier-chasing accounts get carried by non-repeatable moments; consistency accounts get carried by the template itself being the product.

Implication: judge a reel template by whether it can hit the same beat structure every time, not by whether any single instance could go viral. If a crossover/fandom hook is available and honest, it can lift a single reel's ceiling — but don't build the core format around hoping for one.

## 3. Voice sets the pace — voiceover-first workflow

Never fit voice to animation. The order is:

1. Write the script.
2. Generate the audio per sentence and measure actual duration (e.g. one file per sentence via TTS, `ffprobe` the runtime).
3. Time animations/beat-starts to the measured sentence durations — not the other way around.

Natural spoken rate is roughly 180 wpm, but pauses between beats bring the *elapsed* rate down to roughly 110-130 wpm across a full reel. If a script runs long, **shorten the script** — do not speed up the voice track. A rushed voice reads as low-quality and undermines the 5th-grade comprehension target below.

## 4. The 5th-grade rule

Optimize for both comprehension and reading speed simultaneously:

- ~40-50 word total script for a short reel.
- ≤6 words per on-screen text element.
- Every key stat, number, or rule line holds on screen ≥2.5s before the scene advances.
- Charts and big numbers carry the story; on-screen text only labels what's already visually obvious — it should never be the only place an idea lives.

If a viewer would need to pause to re-read a card, cut it down or extend its hold.

## 5. Duration

- Silent (voiceless) reels: 14-20s, biased toward 15-19s.
- Voiced reels: ~22-26s (measured hold times push voiced reels a few seconds past the silent target — that's expected, not scope creep).
- One insight per reel. Don't stack multiple claims into one video hoping for more value; a richer source (e.g. a full blog post) becomes *several* reels, each with its own single insight.

## 6. Data integrity

Every spoken or shown number must trace back to a real data source — no invented stats, no "roughly right" placeholders left in for polish later. When a number is rounded for spoken delivery, say "about" or let the rounding be obviously approximate (e.g. "twenty-eight percent" for a true 27.9%) — never present a rounded number as exact. Keep a written trace from script number → source field/table for every stat that appears, spoken or on-screen.

## 7. TTS craft — phonetic respelling map

Off-the-shelf TTS mispronounces proper nouns, brand names, and acronyms by reading them literally. Maintain a phonetic respelling map (proper nouns, brand names like "TCGplayer" → "T C G player", acronyms like "SIR" → "S I R", spoken letter-by-letter for suffixes like "ex" → "E X"):

- Apply respellings to the **spoken** script text fed to the TTS engine only. On-screen captions always keep the correct/official spelling — never show the respelling to the viewer.
- Generate and sanity-check new respellings individually before using them in a full script (synthesize the single term, check the resulting duration is plausible for its syllable count).
- Produce **per-sentence audio files**, not one monolithic voice track. This makes fixing a single mispronounced line cheap — regenerate that one file and re-splice, instead of re-rendering the whole VO.
- Keep the map itself as a living document per project (term → respelling → note) so it compounds across reels instead of being re-derived each time.

## 8. Visual identity

- **Dark, near-black background** is the default for data reels: it maximizes contrast against data colors (chart lines, highlight colors, ticking numbers), suits night/OLED viewing, makes hero product imagery pop against the frame, and reads as a "terminal aesthetic" that signals technical trustworthiness for a data-driven brand.
- **Hero product/subject imagery** (the actual card, chart, product, etc.) should be sourced at the highest resolution obtainable — check what a source CDN actually serves (don't assume a legacy asset pattern still works; verify with a resolution/HTTP-status check before building around it) and fall back gracefully if a preferred size 403s.
- **Consistent brand typography**, with **tabular figures** (fixed-width digits) for any number that ticks or animates — proportional digits cause visual jitter as a number counts up or down.

From Evan-Kim2028/agent-skills, MIT.

More from Evan-Kim2028/agent-skills