← All guidance

Prompting GLM-5.2

How to prompt Z.ai's GLM-5.2, the leading open-weight model for agentic and coding work (MIT-licensed, ~750B-parameter MoE, 1M-token context). Read the shared Prompting guides principles first; this guide carries only what is distinctive to GLM-5.2. This is reference for when you target, self-host, or evaluate GLM-5.2.

When to reach for it

GLM-5.2 is the open-weight choice for agentic software engineering and long-horizon coding, where it is competitive with the frontier closed models on SWE-bench-style tasks while being self-hostable under a permissive licence. Reach for it when licence, cost, or on-prem control rule out a closed model — or as the open-weight baseline in a cross-model evaluation.

Prompting principles

• Keep tool output in context. On agentic runs, preserve prior tool calls and their results across turns — GLM-5.2 uses them for reasoning continuity and to keep cache hit rates high. Don't strip the trace between steps.
• Control thinking per turn. Thinking is turn-level. Disable it for lightweight tasks (reasoning={"enabled": false}, or clear_thinking: false in content) to get a direct answer; keep it on for hard, multi-step work.
• Escalate effort for hard coding. Switch to deeper/Max compute for complex, multi-step engineering — it produces markedly better results there. Reserve lower effort for simple queries, documentation, and review. In Claude Code the /effort levels (xhigh, max, ultracode) map onto its Max effort mode.
• Structure the prompt explicitly. Reduce ambiguity with a consistent skeleton — goal, context, constraints, inputs, output format, success criteria — rather than an aspirational free-text ask.
• Be concrete on coding specifics. Name the language, framework, and versions, the project structure, and the exact deliverable. For agentic work, give the tools available and a rough decomposition.