Lesson Engine

1 What should it teach?

Write it the way you would say it to a colleague. You are describing a lesson, not writing instructions for a computer — the planner decides the scenes, the narration and what gets drawn.

Topic One or two sentences works best. Naming what you want to see ("show the curve filling in") gives better animation than naming a subject alone.

Or start from an example Each one loads a topic, audience and style that are known to work. Clicking one replaces whatever is in the box above.

2 Who is it for?

These two settings change how the picture looks — text size, colours, how much is on screen at once. They do not change what is taught.

Audience Younger audiences get bigger type and fewer things on screen.

Visual style Set this only when the register differs from the audience — a medical look for a corporate class, say.

3 How long, and how good?

Length Seconds of narration. Longer lessons take proportionally longer to make.

Picture quality Higher quality only changes rendering time, never the content. Draft is right while you are still trying ideas.

Narrator voice SuperTonic runs on this machine and is roughly a third quicker.

4 Which model writes the lesson?

This is the only setting that changes the quality of thinking. Everything else in the pipeline stays identical, which is what makes comparing models fair. Credentials never leave the server.

Provider Local runs on our own hardware — no data leaves the building.

Model Bigger models plan better and take longer. See Compare models for measured differences.

Work continues on the server if you close this tab. Cancelling stops it after the scene it is currently drawing.

Result

Nothing running. Press Generate lesson and the five steps below will report progress here.

idle

PlanThe model writes the lesson as structured data — scenes, narration, what to draw. Checked before anything is rendered; rejected plans are retried.

Find picturesAny diagram the lesson asks for is searched for under a licence we can use, then cleaned. If none exists, a stand-in is drawn and you are told.

NarrateSpeech is synthesised first, so the animation can be timed to the real audio rather than guessed.

DrawEach scene is rendered separately. If one fails, only that scene is affected — a simplified version is used and the shortfall is reported.

CheckThe finished video is inspected for blank frames, missing objects and content errors before you see it.

The lesson plan the model wrote

The animation code this produced

What the checks below mean

Policy	The audience and visual register actually applied — sizes, palette and density come from here.
Assets	Pictures the lesson asked for. RESOLVED means a licensed file was found and embedded. FALLBACK means nothing suitable existed and something was drawn instead — the reason is always given.
Content checks	Whether the lesson is right: formulas that would render as nonsense, a spoken number disagreeing with what is drawn, identical arrows, unlabelled diagrams.
Picture checks	Whether content exists: ink on the frame and every object the plan asked for. This cannot tell you a tidy diagram teaches the wrong thing — watch the video for that.

Lesson Engine — turns a sentence into a narrated animation

Sign in