We read the published research on how people learn from video and turned the findings that survived scrutiny into numeric rules our generator must obey. Here is the whole spec, citations and all, with the things we could not verify clearly marked. It is public because a standard you cannot inspect is not a standard.
Most explainer tools claim to be built on science and stop there. We grade every claim so you can tell a verified effect from a working assumption. The honesty is the credibility.
What the science actually says, with effect sizes and a link to the primary source for each.
Content built up stroke by stroke measurably beats showing completed scenes. Retention ηp²=0.13 to 0.17, transfer ηp²=0.10 to 0.16, replicated across two experiments with identical narration.
In Sketchie: Never cut to a completed diagram. Every element draws on. This is not aesthetic, it is the retention mechanism.
Source: PMC9898452Hand-drawing vs hand-pushing vs no-hand changed intrinsic motivation (p=0.033) but produced no difference in learning outcomes.
In Sketchie: We skip rendering a hand and lose nothing measurable. The draw-on itself is what matters.
Source: Smart Learn. Env. 2023Meta-analysis, 32 studies, N≈3,597: cueing gives retention d=0.27, transfer d=0.34, load d=−0.11. The magnitude of the load reduction predicts the learning gain (β=−0.70 retention, −0.60 transfer).
In Sketchie: Arrows, color pops, and label reveals are cues, not decoration. Every cue points at what the narration is saying right now.
Source: PMC5576760A segmenting meta-analysis found significant retention and transfer gains, small to medium. Part of the benefit comes from the processing time boundaries create.
In Sketchie: Scenes are segments. After a scene’s last reveal, the narration holds silent for at least 0.7 seconds before the cut.
Source: Rey et al. 2019Across 6.9 million edX sessions, engagement sits near 1.0 under 6 minutes, drops to about 0.55 at 9 to 12 minutes, and falls to roughly 0.2 past 12 minutes. Independently echoed in a 2026 replication.
In Sketchie: Default output is 6:00 hard. Longer asks become a chaptered series of videos, never one long video.
Source: Guo et al. 2014About 0.72 vs 0.52 normalized engagement at 3 to 6 minutes, attributed to hand-drawn motion plus an extemporaneous, spoken delivery.
In Sketchie: The format is validated. Narration should sound spoken, not read. Contractions are allowed, direct address is encouraged.
Source: Guo et al. 2014A narrative script improved transfer (ηp²=0.12 to 0.20) but not retention on its own. Paired WITH progressive drawing, narrative beat informative for retention (p=0.021). Over STATIC visuals, an informative script won instead (η²=0.21 to 0.24).
In Sketchie: Our scripts use story grammar, an actor with a goal, an obstacle, and a resolution, precisely because we always render progressively. Never narrative voiceover over a static frame.
Source: PMC9898452Interesting-but-extraneous content does not improve learning.
In Sketchie: Every visual element must be referenced by the narration. Unreferenced decoration is a failure, not flair.
Source: Mayer 2020Prompts that make the learner summarize or explain improve learning. This is the generative activity principle.
In Sketchie: Every video ends with a retrieval beat: a question appears, the video pauses for 1.5 to 2 seconds, then a one-line recap. Nobody in the category does this.
Source: Springer 2020More visual and narrative change points mean less mind-wandering.
In Sketchie: Scene changes are attention resets. Place them at idea boundaries, never mid-idea.
Source: Nature HSSC 2026Explaining quality does not correlate with views once channel size is controlled (r=0.23, not significant). Only likes, relevant comments, and interactions correlate.
In Sketchie: We grade learning design, not virality signals. This was the founder’s instinct, now confirmed with a p-value.
Source: arXiv 2207.05872The laws above become concrete targets. This is the contract the generator is checked against on every render.
| Parameter | Target | Grade |
|---|---|---|
| Total video length | 6:00 hard ceiling; 1:00 to 3:00 sweet spot per concept | Verified |
| Length granularity | 30-second increments | Product |
| Scene duration | 12 to 30 seconds, one idea per scene | Verified |
| Elements per scene | 3 to 7 visible at scene end, heading included | Heuristic |
| Reveal cadence | a new element every 2 to 5 seconds; no two reveals closer than 1.2s | Derived |
| Reveal-to-word sync | element appears within ±500ms of its trigger word | Derived |
| Word-sync coverage | at least 95% of elements carry a reveal word that resolves | Pipeline |
| Post-scene gap | at least 0.7s of narration silence after the final reveal, before the cut | Verified |
| Reveal style | outline draw-on (0.3 to 0.5s), then a staggered fill wash; arrows tip to tail | Verified |
| Visible hand | none | Verified |
| Labels | three words or fewer, all-caps keyword; never a sentence that mirrors narration | Heuristic |
| Narration pace | 140 to 160 words per minute, spoken register, contractions allowed | Heuristic |
| Script structure | hook within 10s, concrete analogy, mechanism, real-world anchor, retrieval beat, recap | Verified |
| Narrative grammar | actor, goal, obstacle, resolution, whenever the domain allows | Verified |
| Analogy | at least one concrete analogy before the first abstract term | Heuristic |
| Extraneous visuals | zero elements unreferenced by narration | Verified |
| Ending | retrieval question, a pause of at least 1.5s, then a one-line recap | Verified |
| Scene transitions | hard cut at an idea boundary only, never mid-idea | Verified |
A rule is only real if you can test it. Every video runs through three layers before it is allowed to ship.
Plain JavaScript assertions on the scene graph and the word timings. Zero cost, safe to run in CI. Total duration under the ceiling, every scene in its duration band, 3 to 7 elements per scene, every element carries a reveal word that lands within ±500ms of the spoken word, a final scene with a question and a recap, and no element unreferenced by narration. Pass or fail, no model calls.
A model-as-judge pass, scored 1 to 5, sampled and run on release, never in CI. One idea per scene, icon fidelity, whether the analogy actually maps to the mechanism, hook quality, a seductive-detail scan, and faithfulness to the source document. Any dimension below 3 fails the release gate.
Checks on the rendered file itself. A contact sheet confirms no overlapping elements and legible labels, a frame burst confirms the outline-then-fill reveal is actually present, and the audio is loudness-normalized with no clipped sentences at the cuts.
The spec is only worth anything if a video that breaks it cannot ship. Layer A runs at generation time, on the scene graph, before we spend a cent on voice or rendering. A graph that fails the mechanical checks is sent back to be regenerated with the exact failures fed into the prompt. A graph that keeps failing never becomes a video.
Every check in the list above maps to a real assertion in our engine: the total-duration ceiling, the per-scene duration band, the element count, the reveal-word resolution rate, the post-scene silence, the label rules, and the question-plus-recap ending. Arrows are checked too, so a pointer never crosses an element it is not connected to. The same code runs in the worker’s regeneration gate and in continuous integration, so a change to the generator that would regress quality blocks itself.
The open eval standard on GitHub Read the Layer-A source Start free trial
