Standalone web application that turns domain knowledge from a Cortex instance into interactive, offline-capable HTML training units, with Venice.ai as the AI provider.
Beta — open source and ready to self-host today, but not yet launched on the Cortex cloud offering. The fastest way to run it: tell your coding agent (Claude Code, Codex, opencode, Hermes, …) to spin up Cortex Trainings and point it at cortexskills.org — it finds full setup context there and helps you connect the app to your Cortex instance. The only thing you provide beyond that is a Venice.ai API key, which accesses and pays for the media generation models (MiniMax H3 and others).
▶ Try a training it produced — a real 7-level course about Cortex itself, generated from a Cortex instance. One HTML file, runs offline.
- Documentation:
docs/— start with docs/README.md - Original service analysis:
OVERVIEW.md - Repo guide for Claude Code:
CLAUDE.md
Briefing ─▶ fable agent ──▶ curriculum.md ──▶ ⛔ approval ──▶ production ──▶ training.html
4 inputs deep research versioned, gate 7 steps single file,
over Cortex reviewable resumable runs offline
(cheap) (costs money)
Part 1 is free, so content gets finished and signed off as a document before any media is generated. Part 2 pauses twice for human judgement: picking the guide-character reference image, and confirming the live video cost quote. See docs/workflow.md.
1. Briefing — four inputs the app refuses to guess (topic, audience, language, duration), plus the visual style for all generated media and an optional collection to scope research to. Optionally, upload up to 3 images of your own guide character and up to 3 style references: a vision model extracts both into prompt text, your character then appears across all films and images (rendered from the actual uploads, keeping its own colors), and everything is generated in the referenced aesthetic instead of the preset style.
2. Research — the agent fans out deep-research and search calls across the knowledge base before it writes a word. Every tool call and result is visible as it happens.
3. Curriculum — a complete document you can actually review: fact sheet, levels, scripts, interactions, cited sources. Revisions are free and versioned; the agent explains what changed.
4. Guide character — production's first pause. Two candidates; you pick the one without baked-in text, because this image anchors the look of every video.
5. Cost gate — every shot is quoted before a cent is spent, and nothing generates until you confirm. Voiceovers are already done by this point (cents); video is the real expense.
6. Done — a single offline HTML file. The step list doubles as the audit trail, and the log shows exactly what was produced and downscaled.
The result of this exact run is live: cortex.eco/demo/cortex-trainings.html.
cp .env.example .env # VENICE_API_KEY, CORTEX_BASE_URL, CORTEX_API_KEY
npm install
npx playwright install chromium
npm run dev # → http://localhost:3000Requirements: Node 22+, ffmpeg and ffprobe on PATH, Playwright Chromium. The Cortex key must
be a plain read-only key (cortex_ro_…), ideally collection-scoped, and the instance needs
ENABLE_AGENTIC_RAG + ENABLE_AGENT_RESEARCH for deep research.
Full reference: docs/configuration.md.
apps/web Next.js app — UI, API routes, agent loop, production pipeline
apps/worker placeholder for extracting production into its own process
packages/shared Cortex + Venice clients, plan/state types
scripts/ qa-training.mjs — automated browser QA of a produced training
docs/ architecture, pipeline, provider notes, troubleshooting
npm run dev # dev server
npm run build # typecheck + production build
npm run typecheck # all workspaces
node scripts/qa-training.mjs <project-id># click through a produced trainingEverything lives on disk under STORAGE_PATH (default apps/web/data) — project metadata,
every curriculum version, the production plan, per-step state, and all media. No database:
the artefacts are the point, so they stay diffable, recoverable, and inspectable. Layout in
docs/architecture.md.
The idea for this comes from Julian Ivanov. His video demonstrated the approach — building a complete interactive learning unit as a single offline HTML file, with the curriculum written and approved before any expensive media is generated — and the skill itself lives in his KI-Automatisierungs-Community.
This repository is an independent implementation of that idea as a web application sourcing its content from a knowledge base. It is not a copy of his skill and uses a different toolchain, so the skill is linked rather than vendored — go to the source for the original. Thanks, Julian.