菜单
Use this skill when you need deep, repeatable multimodal inputs/outputs research across MetaHuman, video generation, and Gaussian splatting, with auto-updated conclusions in deep-research-multimodal_inputs_outputs.md.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use this skill when you need deep verification of slide-5 interaction claims, continuous ArXiv/GitHub research on real-time social avatars, and self-evolving architecture guidance across MetaHuman, video generation, and Gaussian splatting.
Use this skill when you need to proofread research-focused slides, validate claims against primary-source links, avoid hallucinated metrics, and persist repeatable evidence-check history for future cycles.
Use this skill when a task asks for deeper MetaHuman understanding from Unreal source, architecture/module/function extraction, continuous MetaHuman documentation refreshes, or cumulative memory across repeated analysis cycles.
Use this skill when you need deep, repeatable research on realtime Gaussian avatars, continuous ArXiv/GitHub tracking, and auto-updated conclusions in deep-research-Realtime-Gaussian Avatars.md.
Use this skill when you need to continuously collect the latest YouTube research/demo videos for realtime Gaussian avatar systems, publish a curated video-wall dataset, and keep a self-evolving memory of queries/channels over repeated cycles.
| name | multimodal-io-research-evolver |
| description | Use this skill when you need deep, repeatable multimodal inputs/outputs research across MetaHuman, video generation, and Gaussian splatting, with auto-updated conclusions in deep-research-multimodal_inputs_outputs.md. |
Run repeatable research cycles to keep multimodal pipeline claims current for the three avatar approaches:
Each cycle:
deep-research-multimodal_inputs_outputs.md,python .claude/skills/multimodal-io-research-evolver/scripts/evolve_multimodal_io_research.py--skip-core-refresh (reuse latest full-modality outputs)--max-arxiv-per-query 12--max-arxiv-total 80--max-github-per-query 12--max-github-total 80--no-publish (skip writing web/public/docs mirrors)Local skill outputs:
.claude/skills/multimodal-io-research-evolver/references/latest-research.json.claude/skills/multimodal-io-research-evolver/references/latest-report.md.claude/skills/multimodal-io-research-evolver/references/progress.md.claude/skills/multimodal-io-research-evolver/references/history/research-*.json.claude/skills/multimodal-io-research-evolver/references/history/report-*.md.claude/skills/multimodal-io-research-evolver/state.json.claude/skills/multimodal-io-research-evolver/events.jsonlPublished mirrors (unless --no-publish):
web/public/docs/multimodal-io-research-latest.mdweb/public/docs/multimodal-io-research-latest.jsonweb/public/docs/multimodal-io-claim-matrix-latest.jsonweb/public/docs/multimodal-io-research-progress.mdPrimary narrative output:
deep-research-multimodal_inputs_outputs.md (upserted conclusion section)Each cycle must:
state.json and prior latest-research.json when present,events.jsonl,references/progress.md,Do not delete history/events unless explicitly requested.