What Zhipu Qingyan Is and How Users Can Utilize Its GLM Assistant and Custom Agents for Daily Tasks

Zhipu Qingyan (智谱清言) is an enterprise-grade conversational AI platform developed by Zhipu AI, built upon the proprietary General Language Model (GLM) architecture. Operating both as an interactive desktop and mobile web assistant and as a no-code agent creation environment, the platform enables individual professionals, creators, and teams to tackle complex daily workflows. At its core, Zhipu Qingyan combines bilingual natural language understanding, real-time web browsing, code execution in an isolated sandbox, multimodal image generation, and a persistent agent framework. By understanding how the foundational GLM assistant operates alongside specialized custom agents, users can transform routine tasks—from document synthesis and quantitative analysis to automated content drafting—into structured, repeatable digital workflows.

Understanding the GLM Assistant Architecture

The default conversational experience in Zhipu Qingyan is driven by the GLM foundation model family (including GLM-4 and multimodal variants like GLM-4V and CogView). Unlike simple text-completion bots, the platform functions as an integrated workbench equipped with dynamic tool routing. When users submit an inquiry, the underlying model determines whether to answer directly from parametric memory, invoke real-time search, execute Python code, or process uploaded media.

Built-in Tool Ecosystem

The standard assistant includes several integrated modules that activate automatically or upon user request:

• Real-Time Web Search: Queries external web sources to retrieve breaking news, current documentation, financial benchmarks, or recent developments, providing verifiable citations alongside generated responses.
• Advanced Data Analysis (Code Sandbox): Runs an isolated Python execution environment capable of loading tabular datasets (CSV, XLSX), performing mathematical operations, executing statistical computations, and rendering downloadable visualization charts.
• Multimodal Perception and Synthesis: Employs vision models to inspect diagrams, handwritten notes, and photographic inputs, while utilizing the CogView engine to produce high-resolution illustrations and graphics directly from descriptive prompts.
• Long-Document Processing: Ingests large files—such as extensive PDF reports, research papers, Word manuscripts, and slide decks—allowing users to extract summaries, compare conflicting viewpoints, or locate buried data points across hundreds of pages.

General Assistant vs. Custom Agents: When to Use Which

While the default GLM assistant handles ad-hoc, open-ended questions, repetitive workplace responsibilities benefit from dedicated custom agents (known in the interface as 智能体). Custom agents encapsulate fixed system instructions, curated reference documents, and specific tool permissions into a persistent assistant that can be reused or shared across teams.

How to Build a Custom Agent in Zhipu Qingyan

Creating a custom agent does not require software development experience or API knowledge. The platform provides a visual configuration studio accessible via desktop browser at Zhipu Qingyan.

Step 1: Access the Agent Center

Log in to Zhipu Qingyan and select the Agent Center (智能体中心) from the navigation panel. Click Create Agent (+创建智能体) to enter the dual-pane workspace, featuring configuration controls on the left and a real-time interactive preview sandbox on the right.

Step 2: Define Name, Persona, and System Instructions

Users can configure agent behavior either through guided conversational prompts or manual form configuration:

1. Agent Profile: Provide a clear name (e.g., "Executive Briefing Editor" or "E-Commerce Copy Reviewer") and a concise description summarizing its operational focus.
2. System Prompt (Instructions): Define the agent's core role, behavioral boundaries, output format, and constraints. A well-structured instruction set should outline:
   • The primary role and intended target audience.
   • Step-by-step processing methodology.
   • Specific style conventions (e.g., professional tone, Markdown tables, bulleted takeaways).
   • Negative constraints (e.g., "Do not extrapolate beyond provided knowledge base documents; state clearly when information is missing").

Step 3: Attach Private Knowledge Bases

To ground the agent in domain-specific materials without fine-tuning models, upload documents into the Knowledge Base (知识库) section:

• Supported Formats: Upload PDF documents, Microsoft Word files, plain text records, and spreadsheets.
• Automated Indexing: The system parses, chunks, and vectorizes uploaded files for Retrieval-Augmented Generation (RAG). When users interact with the agent, relevant excerpts are retrieved and injected into the prompt context to ensure factual accuracy.
• Best Practice: Clean documents before uploading by removing redundant cover pages, corrupt formatting, or obsolete versions to improve vector retrieval precision.

Step 4: Configure Tools and External Capabilities

Toggle the specific capabilities the agent requires to complete its assignments:

• Enable Web Search if the agent must check live sources or verify external claims.
• Enable Code Sandbox if the agent processes numeric calculations, spreadsheet transformations, or chart generation.
• Enable AI Image Generation if the agent produces marketing visuals or concept illustrations.
• For technical teams, configure Custom Actions (APIs) to allow the agent to fetch data from internal CRM endpoints or enterprise inventory databases via standardized OpenAPI schemas.

Step 5: Test in the Sandbox and Publish

Use the right-hand preview panel to test edge cases, probe negative constraints, and inspect document retrieval quality. Once verified, choose a publication scope:

• Private: Accessible only to the creator's account.
• Link Shared: Usable by colleagues or collaborators via direct URL.
• Public: Listed in the public Agent Plaza for community discovery.

Practical Workflows for Daily Tasks

Deploying Zhipu Qingyan effectively involves pairing the right interface mode with specific operational bottlenecks.

1. Research Synthesis and Long-Document Audits

When analyzing annual reports, academic preprints, or competitor whitepapers:

1. Upload the target document directly into the GLM assistant or a specialized research agent.
2. Direct the model to generate a hierarchical executive summary focusing on metrics, methodology, and strategic risks.
3. Use prompt queries such as: "Compare the capital expenditure figures mentioned in Section 3 against the revenue projections in Section 6, and present the variance in a table."
4. Request explicit page references or section headings to audit the source material efficiently.

2. Quantitative Data Cleaning and Visualization

For weekly metrics reporting or operational spreadsheets:

1. Upload a CSV or Excel spreadsheet into a chat session with the Code Sandbox tool enabled.
2. Prompt the assistant to inspect the dataset schema, detect missing values, and calculate summary statistics.
3. Request specific visual transformations: "Filter rows where conversion rate is below 2.5%, group by marketing channel, and generate a horizontal bar chart displaying average acquisition cost."
4. Download both the generated chart graphic and the modified spreadsheet file produced by the Python sandbox.

3. Enforcing Brand Guidelines in Marketing Copywriting

Marketing teams often spend considerable time editing copy to fit brand voice. A custom agent solves this by standardizing constraints:

1. Configure an agent with instructions specifying brand tone, prohibited buzzwords, character count ceilings, and audience personas.
2. Upload company style guides and product catalogs into the agent's knowledge base.
3. Team members can provide raw product bullet points and receive polished social media posts, newsletter snippets, or advertisement variations that adhere strictly to brand criteria without manual revisions.

4. Meeting Notes Structuring and Action Item Tracking

Transforming unstructured meeting transcripts or rough brainstorming notes into actionable deliverables:

1. Paste meeting audio transcripts or messy bullet points into the assistant.
2. Prompt the model with a structured template: "Extract three sections: Key Decisions Made, Open Questions, and an Action Items table containing Task, Owner, and Deadline."
3. The model eliminates conversational filler, resolves ambiguous references based on context, and produces a finalized memo ready for distribution.

Strategic Tips for Getting Better Results

To achieve consistent output quality across both the default assistant and custom agents, apply the following operational practices:

• Separate Context from Task: Provide background context, source excerpts, and instructions in clearly delineated blocks rather than mixing them into a single run-on sentence.
• Establish Negative Boundaries: Explicitly inform the assistant what not to do, such as avoiding speculative estimates, omitting generic introductory filler, or restricting outputs to specific formats.
• Iterate on Knowledge Chunks: If a custom agent fails to surface specific knowledge base details, re-upload the material as smaller, focused modular files rather than a single monolithic document.
• Leverage Conversational Refinement: Treat initial outputs as drafts. Follow up with targeted adjustments—such as requesting deeper technical depth, reformatting into Markdown tables, or shifting tone—to hone the final result.