glm

name	glm
description	Manually consult GLM-5.2 (Z.ai) via the `zai` CLI — fast by default, `--deep` for reasoning. The user-invoked `/zai:glm` surface; complementary to Claude, never a replacement. For autonomous cross-checks use the `ask-glm` skill instead.
argument-hint	[--deep] [--effort max] [--file PATH] <question>
allowed-tools	Bash
disable-model-invocation	true

glm — manual GLM-5.2 consult (/zai:glm)

Self-Evolving Skill: This skill improves through use. If the zai invocation drifted, a flag changed, or the passthrough broke, fix this file immediately — don't defer. Only update for real, reproducible issues.

Run the zai CLI to consult GLM-5.2 with the user's arguments, then relay the answer (note it's GLM's, not Claude's).

zai chat $ARGUMENTS

• If the user passed --deep/--effort, keep them (extended reasoning). Otherwise it runs fast.
• For a big document, they can pass --file <path> (GLM-5.2 accepts ~1M input tokens).
• For web research, use zai websearch "<query>" / zai read <url> instead.
• Full capability surface: the zai plugin's references/CAPABILITIES.md.

This is the manual surface (disable-model-invocation: true) — it fires only when the user types /zai:glm. When Claude should consult GLM on its own initiative, that's the ask-glm skill, not this one.

Post-Execution Reflection

After running this skill, reflect before closing the task:

1. Did the zai call fail or a flag not behave as documented? Fix the command or the argument-hint here immediately, and note it in the zai plugin's references/CAPABILITIES.md.
2. Did the user reach for /zai:glm when ask-glm or zai websearch fit better? Sharpen this skill's description so the right surface triggers next time.

Do NOT defer. The next invocation inherits whatever you leave behind.