Six open-source pipelines that run inside the AI agent you already use. Feed in a video, a voice note, or an episode script — get back a teacher's guide, ready homework, or a finished lesson in Edvibe.
One 3-minute video becomes Thursday's B2 lesson in your Edvibe course. The same job, done twice — first by hand, then with Prism. (Times are illustrative.)
Open the video. Watch it with a notepad, then rewind the parts worth quoting.
Pull the vocabulary by hand and second-guess the level.
Write four exercises, then add two more so the homework isn't thin.
Format all of it so it pastes into Edvibe without breaking.
Notice the answer key is still sitting in the student copy. Fix it.
Paste the video link and the level: “B2 guide, four plus two, under 15 minutes.”
Read the guide. Drop one exercise that is off for this group.
Paste it into Edvibe and check the answer key is separate.
Done. Ten minutes, start to finish.
105 minutes became 10 — and the guide at the end is the guide you would have written.
Prism is a set of recipes for the AI assistant you already talk to — Claude, Codex, Hermes, Cursor. You send one message. It does the prep.
You install it once with a single line. After that there is nothing new to open or learn — you just describe the lesson you want, in words.
Open source (MIT), no subscription, no server. Your Edvibe login and your students' work stay on your computer — nothing is shared between teachers.
Every artifact is built for one student at one level, with vocabulary taken from the real source. The level never drifts without you asking.
Every artifact below came out of these pipelines — from a real video, a real voice note, a real student. Open any of them; the full examples ship with the suite.
A simulation — click through what each pipeline hands back, then paste the same line into your own assistant.
Each one takes raw material and returns a finished artifact. Press “Try it” to watch any of them run — docs for all six live in the repo.
All six pipelines run the same three steps — and none of them writes a word before the source is in hand.
The pipeline pulls the raw material first: a transcript, a voice note, a script, the text of an exercise. Nothing gets written until the source is in hand.
transcript · script · audio → structured sourceA fixed structure plus level-aware prompts build the artifact in your house style: vocabulary taken from the source, the same exercise layout every time. Quoted lines get checked against the source before anything ships.
source + level contract → your house styleIt ships as a file you can paste, a summary you can send, or a lesson written straight into your own Edvibe account through a local CLI.
artifact → Edvibe · Notion · your studentA stylized peek at the build step: the real command validates a small lesson spec in a dry run, then builds the lesson section by section inside your own account.
Nothing in the middle. Everything runs on your own machine, under your own Edvibe login. The password is typed once, in your terminal, and never leaves it; the session stays cached on your computer, and nothing is shared between teachers. Quoted lines are checked against the source programmatically — nothing reaches a student before you have read it.
Node 18+ and Python 3.9+ on Linux or macOS. Nothing else. No account, no server.
Node 18+ and Python 3.9+. Run node -v and you are half done.
Installs all three skills into every agent it detects. Add -a hermes-agent, -a claude-code or -a codex to target one.
Ask your agent to “connect Edvibe”. It installs the local CLI, walks you through a one-time login, then lessons build straight from chat.
Two commands and one login. MIT licensed, and it runs on your machine.