Style the boards yourself. The included boards are starter templates, not a finished visual identity. Customize the colors, fonts, shapes, and layouts for your own videos. The animation examples are there to adapt.
An open-source workflow for recording, AI-assisted editing, design, Claude Design, animation, motion graphics, and the final editor handoff.
Want more open-source AI filmmaking workflows?
Join the TEIN AI waitlist for updates on new workflows, tools, and releases.
Join the TEIN AI waitlist →
Eyeline connects the script you read, the takes you record, the source clips you cut, and the visuals you design. The workflow keeps its decisions in readable Markdown and JSON so you can use your own AI agent, change the rules, and keep editing by hand.
Quick start · Workflow · Design & motion · JSON formats · Customize the skills · Architecture
| Part | What it does |
|---|---|
| Eyeline teleprompter | Read a script, follow your voice, record microphone audio, mark flubs, track attempts, and optionally sync a display over your local network. |
| Eyeline Edit | Review source-linked clips, inspect waveforms and word timings, adjust cut edges, relink camera originals, and export video/audio. |
| AI editing tools | Transcribe takes, estimate camera sync and drift, build full-take proxies, propose a candidate cut, and report possible transcription or edge issues. |
| Agent instructions | Separate recording, cutting, review, design, animation, and editor handoff responsibilities. Use with a coding agent that can read files and run commands. |
| Design / Claude Design workflow | 267 restyled HTML reference boards (246 animated), eight simple static starters, three filled design briefs, and prompts for planning, building, and reviewing boards. Use your design tool of choice, including Claude Design. |
| Animation / motion graphics | An animated board library with play/pause/scrubbing, JSON timelines, and a local frame renderer for narration-aligned graphics. |
| Editor handoff | A predictable folder tree, approved audio, source mapping, rendered boards, captions, and a finishing checklist for Resolve, Premiere, or another editor. |
| Placeholder skills | Fresh, intentionally unconfigured templates you can fine-tune for your own voice, design, pacing, and motion. No private production skills are included. |
This is an early public release, not an unattended one-click editor. Cutting decisions need human review. Speech recognition can turn a mispronounced word into a correctly spelled transcript, and waveform silence is not proof that a word is finished. Read the limits.
Install Python 3.10+ and uv from their official installers. Chrome or Edge is recommended for browser speech recognition. No API key is required to launch either app.
git clone https://github.com/sirioberati/YouTube-AI-Video-Editing-System.git
cd YouTube-AI-Video-Editing-System
# Teleprompter — open http://localhost:5317
python3 apps/teleprompter/server.pyOn Windows, use py in place of python3. Projects are saved to ~/Eyeline/Projects by default. Set EYELINE_DATA before starting to use another drive. Keep the terminal running.
In a second terminal:
# Editor — open http://localhost:5320
uv run apps/eyeline-edit/edit_server.pyThe editor needs an Eyeline project with a cut plan and full-take playback proxies. For a first look without recording anything:
uv run tools/create_demo.py
uv run apps/eyeline-edit/edit_server.py .local/demo/01-recordingThe demo creates synthetic camera/audio media locally; it contains no real recordings. See setup for Windows, external drives, dependencies, and LAN display mode.
flowchart LR A[Script + Eyeline teleprompter] --> B[Takes + audio + JSON] B --> C[AI candidate cut] C --> D[Eyeline Edit + human review] D --> E[Approved WAV + transcript] E --> F[Design brief + Claude Design / HTML boards] F --> G[Animation + motion graphics] G --> H[Editor handoff + final video]
The recording folder keeps microphone audio, script snapshots, take metadata, captions, and a readable take report together.
Edit the original source intervals. Keep complete source handles available. A playback proxy covers a whole take; an assembled export never replaces the source clips. The experimental assembler writes plan.candidate.json, leaving your active plan.json alone.
Work from the approved final-cut transcript. The unbranded design pack includes:
- 267 restyled HTML reference boards: 246 animated examples and 21 static/reference layouts, preserving the original motion mechanics. Includes typography, charts, diagrams, reveals, masks, transitions, and motion studies.
- Eight additional simple static starters: process, comparison, chart, timeline, layers, code, evidence, and recap.
- Three filled example briefs: a review process, a comparison, and a measurement explanation.
- Three reusable prompts: transcript to brief, brief to HTML, and board review.
- A searchable player with Play/Pause, Replay, and scrubbing, plus rendering examples.
Download or clone the repo, then run python3 -m http.server 5331 --directory templates/design from its root and open http://localhost:5331/ to browse and play the boards. GitHub displays HTML source rather than running it. The library uses a generic editorial theme and system fonts. Its animation timing and effects are preserved, while personal media is replaced with neutral placeholders. Style the boards yourself before production. Example copy, quotes, and numbers require verification or replacement. See the library guide for dimensions, media/rendering limits, and asset credits.
Start with the design brief and Claude Design prompt, or adapt one of the filled briefs. The reference HTML is included with personal branding and signature styling removed. Private skills and identifiable source media are not bundled.
Align graphic changes to the approved narration. Each HTML board exposes window.seek(seconds) so preview and export use the same deterministic state. Start with the generic board, customize your visual system, and render from a timeline JSON. See design and animation.
Deliver the approved cut, its WAV and transcript, source-linked plan, design assets, board renders, and timing manifest. The final editor assembles them in Resolve, Premiere, or another NLE. This release provides the files and handoff contract; it does not automate those commercial editors.
The following diagrams show the example workflow behind the project. Their take counts and thresholds are illustrative, not quality guarantees. The public cut generator proposes a candidate for review.
See the teleprompter layers, editing layers, and example take-selection breakdown
Editable source mapping is preserved. Redundant splits between continuous source intervals are not required. QA remains a combination of diagnostics and human listening.
apps/ Teleprompter and Eyeline Edit source
agents/ Generic role instructions and handoffs
skills/ Placeholder SKILL.md files only
config/ Workflow stages and example preferences
schemas/ Core JSON contracts
examples/ Synthetic JSON examples (no production projects)
templates/ Project tree, design pack, briefs, prompts, HTML boards
tools/ Project scaffolding, demo, render, release checks
docs/ Setup, architecture, data formats, media and guides
The screenshot illustrates a real project's organization. Its project contents and backups are not bundled. Generate your own complete working folder:
python3 tools/init_project.py "/path/to/My Video"Everything in skills/ is a placeholder, written for this public repository. Fill in your audience, voice, editorial rules, design tokens, animation rules, and examples. Customizing these Markdown files is prompt configuration; it does not train a model. Private channel instructions, voice fingerprints, research libraries, and proprietary skill implementations are excluded.
The apps run locally. Browser speech recognition may use your browser vendor's service; QR code and font resources load from public CDNs. Whisper models and Python dependencies download on first use. Sending a transcript to an AI agent or Claude Design follows that tool's data policy. LAN display mode is opt-in and intended only for a trusted private network. See privacy.
License: code and newly authored templates are MIT. TEIN branding is excluded from the MIT grant; its owner retains all rights. Supplied demonstration images and GIFs are documented separately in media credits. Third-party software and model weights retain their own licenses.
Get updates on new AI filmmaking workflows, tools, and releases.