AI Video Factory

English · Türkçe

An AI video factory that produces multilingual, animation-heavy YouTube videos from a single topic. Claude agents write the story and script, plan the shots, design the vector art and review their own output. Paid media APIs generate voice, images and video clips, and Remotion renders the final cut. A web Studio shows what is happening at every step and lets a human make the final call.

Architecture, provider choices and the roadmap are in docs/PROJE_PLANI.md (Turkish). Release history: CHANGELOG.md.

WrongStack: a free, open-source (MIT) AI coding agent that gets better at your codebase over time. It reads code, runs tools and coordinates specialist agents, with durable memory across sessions, visible permission boundaries and no subscription. One self-contained binary, no Node.js needed. macOS / Linux: curl -fsSL https://wrongstack.com/install.sh | sh Windows: irm https://wrongstack.com/install.ps1 | iex wrongstack.com · GitHub

• Topic to finished video. The pipeline runs these stages: style → story → script → approval → cast → concept → approval → voice → shots → keyframes → clips → timeline → render → package → short → review. Every stage is idempotent and saves after each sub-step, so a failed run resumes where it stopped.
• Four motion modes:
  • vector: Claude-drawn SVG characters on a rig, animated in code. This is the cheapest mode.
  • still: images with Ken Burns motion.
  • mixed: some shots become AI video clips.
  • animated: every shot becomes an AI video clip.
• Quality gate. No SVG is used without a check:
  1. Before drawing, a planner agent studies reference images and turns the description into a concrete drawing plan.
  2. A blind reviewer says what the drawing shows without being told its name.
  3. A gate agent scores how recognisable it is and how faithful it is to the references.
  4. Drawings that fail are redrawn with the list of defects. If they still fail, they are not used. A key prop that appears in several shots gets one more try from a simpler, iconic brief; if that fails too, the scenes that used it are replanned without it instead of showing an empty room.
• Grounded in reality. Real subjects are drawn from reference images found on Wikipedia and Wikimedia Commons. For a real person, the photos are first measured once: face shape, brows, eyes, nose, hair and hairline, facial hair, glasses and build go on a likeness sheet (what differs from average, with colours as hex); the illustrator draws to it and the quality gate checks for it. The analysis never names the person or infers anything beyond form and colour, and a person who may be a minor is skipped. For objects, a ready-made public-domain/CC0 vector drawing from Commons is tried first; it is used only if it passes the same quality gate. Claude does not draw maps or flags: they are generated from real data (world-atlas, flag-icons). A map of a historical event shows that year's borders, from the Cliopatria dataset (Seshat Global History Databank, CC BY 4.0, downloaded once on first use), never today's countries; the credit line is added to the publishing description automatically.
• The human has the final say. In the Studio you can accept a drawing the gate rejected, or reject one it passed and add a note for the designer. You can also turn the reality check off for one episode.
• 64 ready-made style packs. Examples include flat explainer, Ottoman miniature, film noir, blueprint, ukiyo-e, synthwave, chalkboard, watercolour, 8-bit pixel, risograph, stained glass, isometric diorama, Byzantine mosaic, Egyptian wall painting, Greek black-figure, Anatolian kilim, ink wash, manga screentone, technical cutaway and whiteboard doodle, plus four 3D-look packs (soft 3D, claymation, vinyl toy, low-poly): in vector mode their characters are real 3D figures built in code with three.js (matte, clay, glossy or faceted surface; a cheap agent reads the approved drawing and picks skin, hair, beard, glasses, clothes and build), they keep every puppet pose, gesture and talking mouth, and they turn to profile while walking; rendering stays free. The concept step, the character bank and the style gallery show the figure itself (a preview rendered locally), marked 3D. Rooms and objects in those packs get their volume from SVG gradients, highlights and contact shadows. In image/video modes the packs ask for real 3D renders. In Automatic mode a cheap agent picks the pack that fits the topic; you can also choose one yourself, or choose Custom to let the art director design freely. A pack fixes the colours, drawing rules and line style, so all episodes in a pack look as if the same hand made them. Run pnpm ys styles to list them. The new-episode form shows a gallery with the same character and room drawn in each pack; click one to choose it. The samples come from episodes already drawn in that pack, and pnpm ys styles samples [pack...] draws the missing ones with the real designer and quality gate (Claude, about $0.2–0.3 per pack). Each sample character also joins the character bank as a presenter in that style. Packs also carry a texture that the renderer lays over every vector scene in code (paper, watercolour, chalk, halftone, engraving hatching, film grain, scanlines), so they differ by more than colour.
• Vector animation that obeys physics. Characters are puppets on a rig, moved entirely in code: things on the floor never float, walk steps follow the distance travelled, held objects stay in the hand, and a character appears only once per shot. Gestures have anticipation and overshoot; pointing aims at its target, and when the object is within reach the elbow bends so the fingertip stops at its edge instead of covering it, eyes follow the listener, and explaining uses one dominant hand. Mouths speak the narration: every sound gets its own mouth shape (lips closed on m, b and p, wide on a, round on o, pursed on u, teeth on the lip for f and v), drawn in code where each drawing's mouth was measured. Eyebrows carry the emotion: each drawing's own brows are split into left and right and tilted in code, inner ends raised when worried or sad, lowered and drawn together when angry, one brow arched while thinking. The head nods, the brows lift and the hands stroke on the stressed words of each sentence. Word timings are aligned to the pauses in the recorded voice, so captions, lips and gestures stop where the narrator pauses. A thinking character takes one of three poses per shot: the chin resting on the hand in front of the face (the elbow propped on the other hand, the fingers kept below the mouth), scratching the side of the head, or arms crossed with the head tilted and the eyes up. Characters also act out what the narration says: typing at a keyboard, writing, reading a page, holding an object up to show it, clapping, hands on hips, arms crossed, handing an object to someone, nodding yes and shaking the head no; clapping, key taps and pen strokes are heard in sync with the hands. Standing characters rest their weight on one leg and shift it now and then, and listeners take a relaxed pose of their own (a hand on the hip, hands together in front) instead of standing in an identical row with arms hanging. Optional side views let walking characters turn to profile. Close-ups reframe themselves when the subject would stay tiny. A small key object in a wide shot (a page, a screen, a chip) is shown magnified in a round inset with a pointer, placed where it covers no face or text. Name tags sit right on the drawn head, the info card moves to the other upper corner when a person is under it, people on one bench sit side by side, two people standing in a wide shot come to talking distance (about 1.5 m) instead of staying at opposite ends of the room, stopping at the edges of a desk or board between them, and a person in front steps aside instead of hiding a face behind them; a monitor on a desk slides along it so it does not hide the face of the person behind. The camera never cuts a standing person at the ankles, shins or knees: either the feet are in frame or the cut is at mid-thigh or higher. A close-up of a person frames them from the waist up with a little headroom (a director's zoom of 1.5–2 used to show a small full figure under an empty wall), and 3D figures are drawn at screen resolution so close-ups stay sharp. When two shots in a row show the same people in the same room at the same size, the second one changes size (wide, medium, close) so the cut does not look like a jump. Confetti, sparks and dust fall behind the characters; nothing covers the on-screen text. Cheering fists pump at head height with bent elbows, each person in their own rhythm, clapping palms meet at the middle of the chest, typing hands rest on the two halves of the keyboard and tap, and the writing hand moves along the line while the other holds the paper. No pose goes beyond what a human body can do: joint limits come from clinical range-of-motion tables (the elbow bends at most 150°, measured in 3D with foreshortened limbs reconstructed; the knee never bends sideways; the shoulder crosses the body only as far as it can), render clamps anything beyond them, and a test sweeps every gesture, every pair of gestures, and sitting, holding, mirrored and walking poses. 3D figures turn a foreshortened forearm or thigh toward the camera instead of squashing it. Before render every shot is checked in code: a half face at the frame edge, a face hidden behind furniture, text over a face, people walking through each other, a hand pointing at nothing, typing away from the desk, a prop that could not be drawn, an empty frame, a joint that jumps between frames, a joint bent past its human limit. What can be fixed without losing meaning is fixed at render: a person steps up to the desk or out from behind the counter, picks up the nearby book to read it, a walker passes in front of someone instead of through them, a cheer lasts a few seconds instead of the whole shot, and the camera keeps every still face fully in or fully out of frame. Scenes that would still show a defect are replanned by the director.
• A vertical Short from every episode. The publishing agent picks the strongest 20–60 seconds of consecutive scenes; code re-renders them as 9:16 at no API cost, with music and effects regenerated for the clip. In vertical frames the stage shrinks per shot when needed so characters at the edges are not cut.
• Music and sound without API costs. Background music is synthesized in code: its sections follow the mood the director gives each shot (bright, warm, cool, night, dramatic), all in one key with crossfades, and it ducks under the narration word by word. In vector mode, animations also get frame-locked sound effects. Both are per-episode switches.
• Clean on-screen text and optional subtitles. Place, date and event lines sit in one info card at the top-left; name tags stay next to their person. Subtitles are always written as an .srt next to the video, and burning them in is a per-episode switch.
• Character bank. Keep topic-independent characters (a presenter, a scientist, a child…) and use them as the on-screen presenter, add them to a cast, or make them a series mascot. They are never redrawn.
• Character reuse. Approved characters are kept in a library together with their identity. When the same person appears again in the same style (for example a real historical figure or the channel's host), the casting agent reuses the drawing instead of redrawing it. In the Studio you can also pick one with Use from library. Run pnpm ys library to list them.
• Series. Episodes can belong to a series. The home page filters by series. A new episode in an existing series inherits that series' settings and style pack, and the series' own characters are offered first for reuse. Use --series "Name" in the CLI.
• Learning loop. After each render the frames are reviewed, and lessons are written to a rule book that later episodes follow.
• Many media models, one router. Paid calls go through one gateway that ranks providers by quality and price for the episode's budget tier, caches by content hash and falls back when a provider fails. Available today: fal (FLUX, Kontext, Seedream 4.5, Kling 3 / 2.6, MiniMax H3 Max Turbo, PixVerse V6, Vidu Q3 Turbo, Seedance 2), Gemini (Nano Banana image models, TTS, Omni Flash video as an opt-in), ElevenLabs (voice, sound effects, music) and a local ComfyUI. Run pnpm ys plan to see the ranking without spending money.
• Cost first. Estimates are shown before any money is spent. Requests are cached by content hash, so the same request is never paid twice. A per-episode budget cap applies to paid image, voice and video APIs. Claude agent calls are shown separately and do not count toward the cap.
• Choose a model per agent. The Studio Models page sets which Claude model each agent uses (for example Sonnet instead of Haiku for the quality gate).
• Bilingual Studio. The Studio is in Turkish and English. Videos can be in any language the voice provider supports. Foreign names and terms are read the way speakers of the narration language say them: before the voice-over, a built-in list and a cheap agent turn them into respellings for the voice (a Turkish narration says "C++" as "si plas plas", not "ce artı artı"), while the captions keep the original spelling and the lips follow the spoken form.

• Node.js 24 or newer, and pnpm 11
• FFmpeg (for the local render test and video clips; if ffmpeg/ffprobe is not on PATH, voice-over durations and loudness use the copy bundled with Remotion)
• Claude: either a signed-in Claude Code installation (subscription) or ANTHROPIC_API_KEY
• Optional API keys: fal, Gemini, ElevenLabs. Providers without a key are disabled.
• Optional: Manim 0.22 and LaTeX for formula and graph boards (see below); without them shots simply have no board.

git clone https://github.com/ersinkoc/AI-Video-Factory.git
cd AI-Video-Factory
pnpm install
cp .env.example .env          # fill in the keys you have; empty providers are disabled
pnpm ys providers             # which providers are active?
pnpm ys render-demo           # no API calls: local Remotion render test

It runs without Docker. Assets go to .data/assets, the cost ledger to .data/ledger.jsonl and videos to .data/episodes. To use Postgres 18, Redis 8 and RustFS (S3) instead:

pnpm infra:up                 # ports: 5433 (pg), 6380 (redis), 9000/9001 (rustfs)
pnpm db:migrate
pnpm test:integration
# .env: STORE=s3, LEDGER=db

In vector episodes the director can put a formula or a real data graph on a board in the middle of a shot. The code turns that recipe into a Manim scene and renders it locally into a transparent clip (no API, no model-written code; LaTeX is limited to math commands). Setup on Windows:

python -m venv .data/tools/manim-venv
.data/tools/manim-venv/Scripts/python -m pip install manim==0.22.1
winget install --id MiKTeX.MiKTeX --scope user
# let MiKTeX fetch missing LaTeX packages on its own:
"%LOCALAPPDATA%/Programs/MiKTeX/miktex/bin/x64/initexmf" --set-config-value=[MPM]AutoInstall=1

Other locations: YS_MANIM (path to the manim executable) and YS_LATEX_BIN (LaTeX bin folder). Boards are cached under library/math.

pnpm studio                   # development: http://localhost:3100
pnpm studio:build && pnpm studio:start   # production build, same port

1. New episode. Click New episode at the start of the top menu; the form has its own full-width page. Enter a topic, then choose the language, length, motion mode, line style, voice, budget, whether the reality check is on and whether subtitles are burned into the video. Each motion-mode card shows its estimated cost; the panel on the right breaks the chosen mode down by stage.
2. Script. Edit the scenes and approve them.
3. Concept. Characters and places are cast and drawn. The quality-gate panel shows each drawing with its references, scores, defects and drawing plan. Accept, reject or redraw with a note.
4. Production. The pipeline rail, live panel and activity log show every agent call with its model, cost and duration. Production stops before the render and waits for your approval: the Shots tab shows a preview frame of every shot, and Approve and render starts the render once you are happy (re-creating shots asks for the approval again; changing only subtitles, sound effects or music re-renders without asking). CLI episodes are approved automatically unless created with --manual (pnpm ys episode approve <id> preview); in vector mode a shot's frame appears as soon as all of its props are drawn. To change part of a vector episode, before or after the render, select shots there (Shift-click for a range), write what should change ("the map should show 1453, with the Ottoman and Byzantine states highlighted") and press Re-create: the review's findings for the selected shots go along automatically, so Select weak shots (score 4 or lower) plus Re-create fixes them in bulk without a note. The director re-plans only those scenes, new props are drawn and the video is rendered again (one director call per scene; the automatic revision passes are skipped so they cannot undo your note).
5. Watch and download the video and its .srt subtitle file, and read its review. Subtitles are always written as an .srt next to the video; burning them in is a per-episode switch (changing it re-renders only, at no cost). The publishing package gives a YouTube title, a description with chapters, tags (each with a copy button) and a downloadable 1280×720 thumbnail. A horizontal episode also gets a vertical Short: the package picks its best 20–60 seconds and they are re-rendered locally as 9:16.

On-screen text follows one layout: place, date and event lines sit in a fixed info card at the top-left, and name tags stay next to their person, all in one typeface and one plate style.

Other pages:

• Costs: the spend ledger.
• Pricing: provider prices.
• Lessons: the learned rule book.
• Rules: the Studio rules, read from the same constants the code uses.
• Models: the model for each agent.

Heavy work runs in a separate CLI process, not in the Studio server, so CLI runs also appear live in the Studio.

pnpm ys episode new --topic "Ottoman coffeehouse bans" --lang en --minutes 3 --motion vector --budget 5
pnpm ys episode run <id>                    # all stages; run again to resume after an error
pnpm ys episode run <id> --until script     # story + script only (no image or voice spend)
pnpm ys episode run <id> --from keyframes   # regenerate from a stage
pnpm ys episode show <id>
pnpm ys cost --episode <id>
pnpm ys plan --cap i2v --duration 5         # provider ranking without spending money
pnpm ys version

Useful options for episode new:

• --motion animated|mixed|still|vector
• --style-pack auto|custom|<pack> (list: pnpm ys styles)
• --line outline|flat|sketch|ink (default: the pack's line style)
• --manual: stop at the approval gates.
• --no-qa: turn off the frame checks.
• --no-realism: turn off the reality check.
• --side-views: vector mode only. Characters that walk get a second, profile drawing and turn to it while walking. Each one costs a designer call and a gate check (about $0.15–0.35) and is kept in the library.
• --vertical: render in 9:16.
• --no-captions: do not burn subtitles into the video (the .srt is still written).

Character bank (vector mode): every approved character is kept in a library and shown on the Studio's Characters page. You can have Claude draw a topic-independent character there (a presenter, a scientist, a child…; about $0.15–0.35 each, gate-checked) and make any character a series mascot. When you start an episode you can pick a bank character as the on-screen presenter (it speaks the narration and appears in most shots) or add bank characters to the cast. They are not redrawn: the episode switches to their style pack. CLI: pnpm ys bank [list|add|mascot], episode new --presenter <id> --with <id,id>.

Run pnpm ys --help for everything.

Keys and switches live in .env (see .env.example). Notable ones:

The version lives in the root package.json. The Studio header shows it, and so does pnpm ys version.

The Studio checks the latest GitHub release every 6 hours. If a newer version exists, a banner with the release notes appears. To update:

git pull && pnpm install && pnpm studio:build   # then restart pnpm studio:start

To cut a release (maintainers):

pnpm release patch|minor|major              # bump, CHANGELOG, commit and tag (local only)
pnpm release minor --publish                # also push and create the GitHub release
pnpm release minor --dry-run                # preview the notes

The release script only runs on main with a clean working tree, after typecheck and tests pass.

MIT © 2026 Ersin KOÇ.

Third-party parts keep their own terms. In particular, Remotion (the renderer) is free for individuals, non-profits and for-profit companies with up to 3 employees; larger companies need a Remotion company license. Generated media is subject to each provider's terms (fal, Google, ElevenLabs), and reference images keep their Wikimedia licenses.