GLM-5V Turbo · workflow
The technical report separates perception, single-step actions, and long-horizon trajectories; this guide focuses on a web-replication loop with tool feedback.
GLM-5V-Turbo, browser and file tools, rollback-capable workspace; target URL and PRD.
Fix the inputs, output contract, and tool boundary; save real returns, errors, and screenshots after each round.
Deliver the artifact for “GLM-5V-Turbo Official Agent Framework Integration and Full-Stack Web Replication Workflow” and list what the inputs cannot confirm.
Run the actual acceptance command and check format, critical paths, and evidence records.
Reproduce the smallest failing case, then narrow the input or fix tool arguments; do not treat model self-report as completion evidence.
The technical report separates perception, single-step actions, and long-horizon trajectories; this guide focuses on a web-replication loop with tool feedback.
By using GLM-5V-Turbo as a vision-language cognitive controller in coordination with agent frameworks such as Claude Code and AutoClaw / OpenClaw, as well as the official skill library (Skills), it is possible to establish a complete loop covering “visual exploration of a target URL and asset retrieval -> full-stack frontend replication,” “PRD-driven web development,” and “multimodal spatiotemporal video analysis.”
Suitable tasks: Autonomous GUI exploration and high-fidelity replication based on a target website URL, PRD-driven Web application generation, multi-page e-commerce / brand-story UI design, second-by-second target tracking in long videos, and 3D spatial bounding-box detection.
Unsuitable tasks: Purely theoretical derivation of complex backend architectures detached from an execution-layer sandbox; or expecting a single model to perform physical operations directly when no external tool executor is configured, such as when file-system read/write or browser automation capabilities are unavailable.
Applicable model version: glm-5v-turbo; supports mixed video, image, text, and document inputs, with a context length of 200K and a maximum output of 128K tokens.
Applicable clients, agents, or APIs: Claude Code, OpenClaw, AutoClaw, ZCode, Z.AI Official Chat Completions API.
Recommended reasoning mode and parameters: The official recommendation is to enable thinking mode with thinking.type="enabled"; for complex full-stack generation and PRD parsing tasks, retain the reasoning process to ensure strict alignment between the visual hierarchy and code structure.
glmv-web-replication skill)Given a target website URL: <TARGET_URL>, please explore it via GUI, collect the necessary assets, and recreate the webpage in HTML code with high visual fidelity and functional completeness.glmv-prd-to-app skill)Based on my PRD document, please use your PRD skills to build a website for the project in the act folder. The working directory is ./act_workspace.You are a master of frontend recreation and web design. Please complete the following design tasks and implement everything in HTML code.
1. Recreate all pages of such a shopping website, using valid image URLs.
2. Create a welcome page and then transition into the shopping interface.
3. On the “About Brand” page, use parallax scrolling to tell the brand story, allowing text to appear rhythmically as the image background moves.
4. Design a color scheme that preserves a premium aesthetic in dark mode and resolves the issue of product images blending into dark backgrounds.
5. Design a one-page checkout interface to reduce user drop-off, including dynamic shipping calculation and address autocomplete.
In addition to the above, also implement all button functionalities, such as Home, Products, About Brand, and Checkout.Please recreate the mobile app interface based on the provided image, and additionally mock several possible follow-up pages or user interactions that fit the same product design and functionality.Based on the description of the objects appearing in the video "<OBJECT_OR_ACTION_DESCRIPTION>", please track the objects corresponding to this description at every second (tracks per second) of the given video, and provide the bounding box and a globally consistent label for each object. Output using valid JSON format.Please locate the 3D bounding box of <TARGET_OBJECT> and output it in JSON format, where the 9 coordinate values correspond to the center point (x, y, z) and the sizes (x_size, y_size, z_size) across three dimensions all in meters, and the three rotation angles in radians.Environment and skill preparation: Introduce the official skills (such as clawhub.ai/zai-org/glmv-web-replication or clawhub.ai/zai-org/glmv-prd-to-app) in OpenClaw / AutoClaw or Claude Code.
Multimodal perception phase: Pass in the target URL, design mockup, or video material. GLM-5V-Turbo serves as the vision-language controller (Vision-Language Controller), calling browser screenshot, cropping, DOM inspection, or OCR tools to extract visible elements and layout constraints.
Planning and code implementation phase: The model performs high-dimensional task decomposition and generates structured code with embedded CSS/JS, or delegates the task to a workspace directory.
Agent execution and visual alignment validation: Claude Code / AutoClaw starts a service in the local sandbox and captures rendered frames to send back to the model, which compares them with the reference image to fix pixel-level and interaction-level discrepancies.
Artifact packaging and delivery: Output runnable frontend project code, a list of assumptions, and visual regression checks.
The official paper positions GLM-5V-Turbo as a native multimodal agent foundation model (Native Foundation Model for Multimodal Agents), integrating the CogViT vision encoder and the Multimodal MTP architecture.
The official release includes a complete official skill list (Table 2), covering glmv-web-replication, glmv-prd-to-app, glmv-pdf-to-web, glmv-pdf-to-ppt, glmv-stock-analyst, and others, and these skills have been open-sourced on ClawHub.
Section 3.2 of the paper explicitly describes the collaboration paradigm with Claude Code and AutoClaw: Claude Code handles terminal and file-system execution, AutoClaw handles browser and GUI automation, and GLM-5V-Turbo handles high-level visual understanding, planning, and orchestration.
Automated replication requires the execution environment to have external network access or local sandbox support; if the target page includes strong anti-crawling measures or CAPTCHAs, an external automation layer is required.
The “valid image URLs” in the official prompt must be replaced with local placeholders or paths downloaded by an asset-extraction script in offline or closed environments.
3D bounding boxes and second-by-second video tracking depend on clear resolution and correct camera-intrinsic assumptions and cannot replace high-precision sensor data such as LiDAR.
The official technical report states: “GLM-5V-Turbo acts as the vision-language controller for AutoClaw, enabling sophisticated agentic workflows... facilitating a complete perception–planning–execution loop.” (arXiv:2604.26752 Section 3.2).
arXiv / Z.AI & Tsinghua University · Source date: 2026-05-12 · Edited: 2026-09-20
Read the original sourceGLM-5V Turbo
Run this guide in the environment listed above. Downloading does not transfer the template or establish model availability for your account.