MartinDelophy/ai-video-editor

Goal Improve the public edit-timeline-studio Agent Skill through reproducible, real-world evaluation and focused fixes. The Skill already supports Codex, Claude Code, Copilot, and Gemini CLI, and can route work through either the versioned command runner or the browser editor. This issue is for improving how reliably an Agent discovers, interprets, executes, verifies, and hands off those workflows—not for adding unrelated editor features. Relevant files: skills/edit-timeline-studio/SKILL.md skills/edit-timeline-studio/agents/openai.yaml skills/edit-timeline-studio/references/ skills/edit-timeline-studio/scripts/validate_edit_plan.mjs skills/edit-timeline-studio/docs/ Issue #37 remains the capability matrix. This issue focuses on Skill quality and behavior. Areas to investigate Contributors may focus on one area: Triggering and discovery: false negatives, false positives, unclear user-intent coverage, or stale UI metadata. Context efficiency: duplicated guidance, unnecessary contex

Summary Translate CONTRIBUTING.md into one language already supported by Timeline Studio so first-time contributors can understand the workflow more easily. Available starting choices include Simplified Chinese, Japanese, Korean, Spanish, French, German, Brazilian Portuguese, Thai, Vietnamese, and Russian. Before starting Comment with the language you would like to translate. Please wait for confirmation so two contributors do not unknowingly work on the same file. Use the filename pattern: CONTRIBUTING.<locale>.md Follow the locale naming already used by the translated README files. Acceptance criteria The translation preserves the meaning and structure of the English guide. Commands, filenames, labels, URLs, and code blocks remain technically accurate. The translated guide links back to the canonical English CONTRIBUTING.md. The main CONTRIBUTING.md gains a link to the new translation. The matching translated README links to the translated contribution guide. The pull request states

Goal Improve the perceptual quality of Timeline Studio's browser-local Digital Human pipeline, especially lip synchronization, identity preservation, temporal stability, and facial motion. The current implementation uses: JoyVASA audio-driven motion in src/workers/joyvasa.worker.js LivePortrait WebGPU rendering in src/workers/liveportrait.worker.js the orchestration flow in src/hooks/useAvatarGeneration.js 256 px preview and 512 px quality modes adaptive neural keyframes interpolated to the output frame rate corrupt-frame detection, bounded retry, and previous-frame hold The models and end-to-end browser pipeline already work. This issue is about making the result look and feel better, not replacing it with a hosted service. Quality problems to investigate mouth movement that is early, late, weak, or unrelated to phonemes unstable lips, eyes, jaw, or head pose between neural keyframes identity drift and changes in face shape or skin tone visible jumps caused by keyframe selection or

Goal Prototype an experimental, browser-local quality gate and repair pass for AI-generated SVG assets. The existing sanitizer correctly blocks unsafe or malformed SVG, but syntactically valid output can still be unusable: artwork may be clipped, nearly empty, dominated by a full-canvas background, poorly grouped, excessively complex, or unrelated to the request. This issue should add a quality layer after parsing and before the asset is accepted. Proposed approach 1. Deterministic quality report Create a pure, testable analyzer that returns structured warnings or error codes for conditions such as: missing or invalid viewBox artwork outside or tightly clipped by the viewBox effectively empty visible content unintended full-canvas opaque background excessive shape count or filter complexity invalid, duplicate, or unhelpful semantic group IDs extreme coordinate values very low color contrast where it can be detected reliably The analyzer must not execute SVG code or rely on a remote s

Goal Improve the visual quality and request fidelity of SVG assets generated locally by Gemini Nano. The current pipeline already detects and translates the request, prompts the browser model, extracts the first SVG document, sanitizes it, and adds the result to My assets. This issue focuses on making the result genuinely useful—not merely valid XML. Relevant code: src/lib/geminiNanoVector.js src/lib/geminiNanoVector.test.js the AI Design flow in src/components/panels.jsx Research questions Which prompt patterns produce the most reliable composition, hierarchy, color, and editability? Which requests commonly produce clipped, generic, overly simple, or visually incoherent output? Do semantic groups, titles, shape limits, or explicit design constraints improve results? How much quality is lost when a non-English request is translated to English before generation? Which quality checks can run deterministically without a remote service? Phase 1 — establish a baseline Create a repro