Docs
How Vidacle works
Vidacle turns a structured intent — video type, category, audience and goal — into one to three ranked, shootable video packages. Each package carries a concept, a script, a scene-by-scene breakdown and a design brief.
01 · Output
What you get back
Every synthesis returns up to three packages, ranked best-first. A package is a complete plan for one video: a concept and its rationale, a word-level script, a scene timeline covering the full duration (visual, audio and text overlay per scene), and a design brief with color palette, typography, format, thumbnail direction and music.
The optional free-text intent field steers the concepts. The structured inputs — type, category, audience, goal — decide which slice of the data your priors come from.
02 · Ranking
How packages are ranked
Vidacle does not predict virality. Packages are ranked relatively — against each other — and every package cites directional priors drawn from aggregates of a large public TikTok dataset.
An LLM judge scores each package on four axes, each out of five, for a total out of twenty:
Hook — does the opening earn the next second?
Clarity — is the idea legible at viewing speed?
Goal fit — does the package serve the goal you picked (saves, shares, likes or reach)?
Prior adherence — does it follow what the data's strongest slice actually does?
The judge writes a one-line reason for its total, shown on each package card and in the package detail.
03 · Reading results
The priors panel
The evidence panel at the top of every run shows the aggregates behind the ranking. How to read each part:
Metric — the rate tied to your goal — save rate for saves, share rate for shares, like rate for likes, views-per-day for reach — with its mean and sample size n.
Fallback level — how close to your exact slice the priors are. Level 0 · exact slice means the aggregates come from precisely your type × category × audience × goal; higher levels mean the slice was widened because it was too small.
Best duration — the duration bucket (15s, 30s, 60s, 180s, long) with the strongest metric in your slice.
Question & CTA lift — how much the metric moves when the caption asks a question or carries a call to action, versus the slice mean.
Top tags & music — the tags and tracks with the strongest lift in the slice, each with its sample size.
Exemplars — real captions from the slice — reference only, never to copy.
Replicated — whether the winning duration, tags and music hold up across a two-fold holdout. A ✕ means the signal didn't replicate; treat it as weak.
Caveats — always surfaced. The most common: the sample is currently a single shard of the dataset, so every number is directional.
Below the panel, the ranked cards preview each package's hook, judge score and rubric. Selecting a card opens the full package: the script with a copy button, the scene timeline with proportional time bars, the design brief's palette swatches, and the judge's four-bar rubric with its reason.
04 · Inputs
Inputs reference
Video type — talking head, voiceover + b-roll, text on screen, skit, tutorial, or product demo.
Category — fitness, cooking, beauty, comedy, dance, education, finance, gaming, pets, travel, fashion or tech.
Goal — saves, shares, likes or reach. There is no "follows" goal: the dataset carries no follower field, so a follows prior cannot be computed honestly.
Audience — a country:language proxy — for example US · English or BR · Portuguese. The dataset has no audience labels, so audience is proxied by the creator's country and caption language.
Format hint — optional — 9:16, 16:9, 1:1, or auto. It nudges the design brief, not the priors.
05 · Limitations
Honest limitations
Priors are directional, not predictive: they describe what a slice of public videos looked like, not what yours will do. The current sample is a single shard of the dataset. Categories are inferred from captions and hashtags, not labeled by the platform. Audiences are a country:language proxy, not a demographic read. Treat every number as a prior to be tested, not a promise.