Zhipu Official: Prompt Writing Guide (GLM Coding Best Practices)

Core content summary

This official prompt guide, “Vibe Coding Basics,” is designed for collaboration between GLM models and coding assistants. It presents the official prompt-engineering methodology for GLM-5.3 (GLM-5.3 is also a flagship coding model in the GLM series and therefore directly applicable).

I. Prompting techniques: How to build rapport with AI

1. From vague to precise

• ❌ Wrong example: Help me write a login feature

• ✅ Correct example (describe what you want + why + how you want it):

I need to create a user login feature for a React application. Requirements:
- Use TypeScript
- Include email and password validation
- Support keeping users signed in
- Make error handling user-friendly
- Follow the project's existing component structure

Please first analyze the existing authentication-related code, then provide a complete implementation plan.

2. The SMART principle: Borrow the SMART principle from project management to structure prompts (specific, measurable, achievable, relevant, and time-bound).

3. Progressive dialogue: from simple to complex — Break complex tasks into multiple steps:

• Step 1, exploration and understanding: Analyze the project architecture and design patterns in the authentication code

• Step 2, design and planning: Design a third-party login architecture based on the analysis

• Step 3, implementation and testing: Implement the feature (frontend components + backend API)

• Step 4, optimization and refinement: Improve error handling and the user experience

II. Guide to using coding assistants

1. GLM models' project-awareness capabilities (GLM can understand the structure of an entire project):

• Practical tip: First ask the AI to “take a look” at the project — Please analyze this project's structure and main technology stack; use existing code as a model — Please follow the style of components/Button.tsx to create a new Card component.

2. Slash commands (an efficiency multiplier), using Claude Code as an example:

• File operations: /create-component, /refactor-function, /add-tests

• Project management: /commit, /pr-review, /fix-issue #123

• Code quality: /optimize, /clean, /security-check

3. Context management (most important for GLM) — Establish a project-context template:

This is a [project type] project that uses [technology stack].
Its main function is [description of core functionality].
Our coding conventions include:
- [Convention 1]
- [Convention 2]
- [Convention 3]

I currently need [specific requirement]. Please provide a solution based on the context above.

III. Advanced techniques

1. Error-driven learning loop — When the code does not behave as expected, provide feedback instead of asking again:

There is a problem with the code you just generated: when the user submits an empty input, it should display a friendly message instead of throwing an exception.
The current behavior is: [describe current behavior]
The expected behavior is: [describe expected behavior]
Please modify the relevant validation logic.

2. Progressive complexity management: Do not start by throwing an overly complex task at the model; follow the principle of “divide and conquer.”

IV. Hands-on exercise: Adding a comment feature to a blog system

1. Establish context (a Next.js + TypeScript blog system with existing article display, authentication, and a Tailwind UI library; add comments, replies, real-time updates, and Markdown support) → first analyze the implementation plan

2. Architecture design (database model + API interface design)

3. Incremental implementation (data model + CRUD API)

4. Frontend components (consistent with the style of the existing components)

5. Optimization and refinement (like/unlike, reporting, and sensitive-word filtering)

V. Common pitfalls and solutions

“Remember: a good prompt is half the battle, and a good feedback loop is the other half.”

Suggestions for combining with GLM-5.3

• With a 1M-token context window and reasoning always enabled, GLM-5.3 can make full use of its long-horizon Agent capabilities through a project-context template + multi-turn progressive dialogue.

• Official prompt-engineering page (platform-level): https://docs.bigmodel.cn/cn/guide/platform/prompt.