Opus 5.5 makes a one-click promo film for the CodePilot app (歸藏 guizang.ai, X video, Chinese)
歸藏(guizang.ai) (@op7418) · 2026-09-24 · ai-made · 179,071 views · 中文
Made by AI
Model: Claude Opus 5.5 · Series: Code-rendered film (LLM writes the program that draws every frame)
Evidence: X post (2026-09-24, Chinese): Opus 5.5 made a promo video for his CodePilot product, 'generated in one click'; he says it beats GPT-6 Astra from a few days earlier and will feed the lessons into his guizang-product-video-skill.
Human role: Creator asked Opus 5.5 for a product promo using his own skill (guizang-product-video-skill), which reads the codebase for components, colors and logos (per his quoted earlier post).
Pipeline: Opus 5.5 + guizang-product-video-skill → front-end (HTML/JS) rendered promo film (50 s)
What's in the video
Description written by Gemini, which watched and listened to the whole video.
Summary
This video is a sleek, Apple-style promotional showcase for CodePilot, a desktop client designed to manage and run multi-model AI agents. Produced autonomously by Claude Opus 5.5 and published by Chinese creator 歸藏 (guizang.ai), the video highlights unified model management, mid-conversation model switching, agent execution safety, and Model Context Protocol (MCP) integrations.
What is shown
- [00:00 - 00:04]: Space warp animation rapidly cycling through model logos (Claude, Gemini, Kimi, DeepSeek) with the caption questioning the constant switching between different model web apps.
- [00:05 - 00:08]: Dynamic logo build and product introduction: CodePilot ("一个桌面，连接所有模型和 Agent").
- [00:08 - 00:13]: Desktop application UI overview under the tagline "All your agents. One window." displaying active agents, session options, and model selection.
- [00:14 - 00:23]: Agent execution workflow ("Say it. It gets done"): user inputs a prompt (
把登录页改成深色主题，再跑一遍测试), and the agent inspectspage.tsx, writes changes, and requests user approval before executingnpm testvia terminal (Allow Once), finishing with 24 passing tests. - [00:24 - 00:29]: Mid-thread model switching ("Switch mid-chat"): switching runtime/model from Anthropic Claude Sonnet 4.6 to Zhipu GLM-5.3 within an ongoing conversation without losing context.
- [00:30 - 00:33]: Provider ecosystem display indicating compatibility with 17+ model providers (including Anthropic, OpenAI, Gemini, DeepSeek, Moonshot/Kimi, Zhipu GLM, MiniMax, Alibaba DashScope, ByteDance Volcano Engine, Xiaomi MiMo, AWS Bedrock, OpenRouter, and Ollama) and multiple runtimes (Claude Code, CodePilot, Codex, Cline, OpenCode).
- [00:34 - 00:40]: Extensibility menu displaying one-click installs for community skills and MCP servers (e.g.,
/frontend-design,/code-review,/humanizer-zh,/find-skills). - [00:41 - 00:44]: Rapid feature checklist over terminal code backgrounds: Split view (分屏双会话), persistent assistant memory (助理长期记忆), and "Built for every day".
- [00:45 - 00:50]: Closing splash screen showing cross-platform availability (macOS, Windows, Linux) and website link (
codepilot.sh).
Claims & numbers
- The video states CodePilot supports 17+ model providers with bring-your-own API keys.
- The interface lists support for multiple runtimes (e.g., Claude Code, CodePilot, Codex).
- The demonstrated terminal test run reports
Tests: 24 passed, 24 total. - Available natively on 3 desktop platforms: macOS, Windows, and Linux.
Notable quotes
- "All your agents. One window. 一个窗口，装下所有 Agent。" [00:10]
- "Say it. It gets done. 说出需求，它自己动手。每一步都看得见，跑命令前先问你。" [00:19]
- "Switch mid-chat. 聊到一半，也能换模型。换服务端和模型，对话接着聊。" [00:26]
Assessment
This is a polished concept-and-feature promotional video made via automated generation and scripted UI animation to demo the CodePilot application. The demonstrated UI interactions (autonomous file editing, CLI confirmation gates, skill marketplaces) showcase realistic workflows of modern agentic coding environments rather than hyper-futuristic CGI.
Lyrics & themes
- Narration/Audio: The video is set to an upbeat, rhythmic electronic instrumental track accompanied by crisp interface sound effects (whooshes, key presses, confirmation chimes) without vocal singing.
- On-Screen Narrative: Follows a classic product launch structure:
- Problem: Fragmented workflow from jumping between disparate AI providers ("今天用 Claude... 明天又换一个?" [00:00]).
- Solution: A single unified workspace ("All your agents. One window" [00:11]).
- Control & Safety: Autonomous execution with human-in-the-loop verification before destructive actions ("跑命令前先问你" [00:20]).
- Extensibility: Unbounded tooling via MCP and skills [00:37].
Lore & references
- Multi-lab agnosticism: Features prominent 2025–2026 Chinese and Western AI models (Claude Sonnet 4.6, Opus 4.7/4.8, GLM-5.3, Kimi, DeepSeek, Xiaomi MiMo).
- Model Context Protocol (MCP): Highlights Anthropic's open MCP standard as a core extensibility mechanism for local developer agents.
- Claude Code runtime: Explicitly lists Anthropic's
Claude Codealongside OpenAI'sCodexand OpenCode as switchable runtimes within the client UI.
Visual style & craft
- Features an ultra-minimalist, sleek dark-mode aesthetic heavily influenced by Apple product launch keynotes (horizon-lit curved sphere, clean typography, dynamic zoom transitions).
- Uses programmatic motion design, simulated terminal typing, and floating macOS window shadows combined with synchronized UI click animations.
Described by gemini-3.8-flash on 2026-10-06 from the video's audio and frames.