GLM prompt guide: constraints early, useful roles and explicit JSON

How to write prompts for GLM-5.3 and GLM-5.3 Flash from Z.ai (Zhipu AI), following Zhipu's general prompt-engineering guide and the GLM-5.3 docs.

Last updated: · Markdown version

What Zhipu recommends

  • Give the important constraints you already know as early as possible. A rule that arrives at the end is easy to under-weight.
  • Roles should serve the task. Skip inflated identities like 'world-class expert'; a role cannot replace the task description, the source evidence or the validation criteria.
  • Name tags after what they contain and keep them consistent, so the model can tell the requirements it must follow from the material it must process.
  • The number of examples is not the goal: cover the differences that matter.
  • For JSON, spell out the structure, what each field means and the allowed values. A structured prompt describes the result but cannot guarantee valid data, so validate it in your app.

What is different about GLM-5.3

  • Reasoning is always on: disabling it is no longer supported. Pick reasoning_effort low, high or max (default max); low is the setting for quick tasks.
  • GLM-5.3 takes text only, with a 1M-token context and up to 128K tokens of output. Use GLM-5.3 Flash for images, video and files.
  • Because it already reasons, you do not need 'think step by step'. Spend the words on constraints and on how to check the answer.

Templates

Constraints-first review

Constraints (apply to everything below):
- Audience: [who]
- Max length: [limit]
- Do not: [what to avoid]

Task: review the draft below and return the 5 most important fixes, most important first.

<draft>
[paste the draft]
</draft>

Done when: each fix names the sentence it applies to and the reason.
Opens the analyzer with the prompt prefilled.

JSON with allowed values

Read the review below and return JSON only:
{"sentiment": "positive" | "neutral" | "negative", "topics": string[] (max 3, lowercase), "refund_requested": boolean}
- sentiment: the reviewer's overall view.
- topics: what the review is about (e.g. "delivery", "price").
- refund_requested: true only if the reviewer asks for money back.

<review>
[paste the review]
</review>
Opens the analyzer with the prompt prefilled.

FAQ

Should I ask GLM-5.3 to think step by step?
No need: it always reasons. Adjust depth with reasoning_effort and use the prompt for constraints, material and success criteria.
Z.ai or Zhipu AI?
Both are Zhipu AI platforms: Z.ai (docs.z.ai) for international users and BigModel (bigmodel.cn) in China. Promptea lists the model ids from the Z.ai docs.