Skip to main content
Home / Portfolio / Studio School
Education

Studio School

A self-hosted course for making AI video on a local GPU and posting it

Built by Rogue AI · 21 lessons, per-person progress and works log · Self-hosted, local lab

Built in November 2026 for two beginners in a local lab who wanted to make their own clips. The first version shipped the course, notes and works log; a lesson on video prompting followed the next day; a week later the checklists became tickable, confidence ratings fed a revisit list, and works gained structured posting numbers.

Studio School, A self-hosted course for making AI video on a local GPU and posting it

The problem

Learning AI video from scratch means scattered tutorials written for other hardware, other model versions and paid cloud tools. What the learners had was a laptop with an 8 GB GPU, local picture and video models, and a set of working scripts, but no path from the first still to a finished, posted clip. They needed a course tied to the tools on that machine, with honest timings and a place to log what they made.

What I built

A course app with 21 lessons in 8 modules across two tracks: Make (15 lessons, from orientation through stills, motion, the cut and other kinds of video) and Put it out there (6 lessons on platforms, hooks, publishing, reading the numbers, rights and money). Each lesson has a goal, a body, an exercise, a tickable checklist and pitfalls, and points at the real scripts. Per person it keeps progress with a 1 to 5 confidence rating, private notes per lesson, and a works log with platform, link and the numbers after 48 hours. A toolkit page collects every command, and a glossary explains the words.

Architecture

Course as code
Modules, lessons and glossary live in one typed content file rendered by a small Markdown component; editing a lesson is a text change with nothing to migrate
Per-person data
PostgreSQL via Prisma 7 with the PrismaPg adapter holds lesson progress, tickable checks, confidence, notes and works, every API call scoped to the signed-in user
Learning loop
Lessons rated 2 or lower appear on a revisit card in course order; works posted 48 or more hours ago without numbers appear on a due-for-numbers card
What the course teaches
The local toolchain on the learners' own machine: ComfyUI on the host GPU, FLUX.2 klein for stills, Wan 2.2 for image-to-video, ffmpeg for the cut, each with the times measured there
Tests that check the course
Unique slugs, ordered lessons, a valid prev/next chain, and every path and script a lesson names checked against the disk; 33 tests in 3 files
Auth and network
Vendored custom JWT core with lockout and rate limits, two seeded accounts, a separate CRUD-only runtime database role, served over HTTPS on the home network only

Tech stack

Next.js 16React 19Prisma 7PostgreSQLTailwind 4TypeScriptVitestCustom JWT (jose)

What broke first

  • ▸

    A course that names real commands rots the moment a script moves. A test walks every lesson and checks that each folder, script and compose file it names exists on disk, so a broken instruction fails the suite instead of failing a learner.

  • ▸

    I keep the course in code, not in the database. Lessons are edited like text and nothing needs a migration; only what belongs to a person, progress, notes and works, lives in PostgreSQL.

  • ▸

    Checklist ticks are stored as the item text, not an index. When I reword or reorder an item in a lesson, an old tick drops off instead of quietly landing on the wrong line.

Outcome

Two beginners have one place that takes them from the first still to a posted clip, with instructions that match their machine and a check that those instructions still run. Progress, notes and the works log are per person, and nothing leaves the computer.

Honest limits

It is a course for two people on one machine, and the lessons are written for that machine: its paths, its GPU and its measured render times. The app renders nothing and posts nothing; it never calls ComfyUI or a language model. Views and likes in the works log are typed in by hand after 48 hours, not pulled from platform APIs.

Related reading

← Back to portfolio