glm5-optimizer

name: glm5-optimizer description: Guidance for optimizing OpenClaw agents running on GLM-5 (z.ai). Use when the user wants to improve GLM-5 response quality, reduce verbose filler, prevent loop failures, or understand GLM-5 behavior. Provides best practices and optimization rules without modifying existing setup.

GLM-5 Optimizer

Guidance for optimizing any OpenClaw agent running on GLM-5 (z.ai) for better performance, reliability, and response quality.

When to Use This Skill

• User asks how to improve GLM-5 performance
• User wants to reduce verbose filler in responses
• User is experiencing loop failures (repeated tool calls)
• User wants to understand GLM-5 strengths and weaknesses
• Setting up a new agent on GLM-5

What This Skill Provides

This skill provides knowledge and guidance, not file replacement. Apply these optimizations to your existing setup:

• Optimization rules to add to your SOUL.md
• Response length guidelines
• Known GLM-5 behaviors (strengths and weaknesses)
• Prompting patterns that work best

Critical GLM-5 Optimizations

1. Loop-Breaker Rule

Add to your system prompt:

If the same tool call fails or produces the same unwanted result 3 times in a row:
1. STOP immediately
2. Try a completely different approach
3. If still stuck, ask the user for help

Why: GLM-5 can "think" about what to do but fail to translate it to tool parameters, causing infinite loops.

2. Action-Reasoning Alignment

Add to your system prompt:

Before submitting any tool call, verify:
- Thinking says "include X" → Check that X is in the parameters
- Stated intent matches actual tool call
- If thinking and action don't match, fix the action

Why: GLM-5's reasoning can disconnect from its actions.

3. Response Length Rules

Add a response length table:

Rule: If response exceeds 100 words, use bullet points or numbered lists.

Why: GLM-5 tends to over-explain by default.

4. No Filler Phrases

Add explicit prohibition:

Never say:
- "Great question!"
- "I'd be happy to help!"
- "Let me assist you with that."
- "I understand."
- "Absolutely!"

Why: GLM-5 defaults to verbose politeness markers.

5. Thinking Block Brevity

Add to your system prompt:

Keep thinking blocks focused on what to do, not detailed execution:
- Complex calculations → Use tools (exec, scripts)
- Long reasoning → Move to tool execution
- Thinking should be: goal → approach → go

Rule: If thinking exceeds 200 tokens, you're overthinking. Act instead.

Why: GLM-5 can generate 1000+ token thinking blocks.

GLM-5 Model Characteristics

Strengths

• Step-by-step reasoning
• Code generation
• Long context handling (205k tokens)
• Tool/function calling
• Multilingual (Chinese/English excellent)

Weaknesses (Addressed by This Skill)

• Reasoning-to-action disconnect (loop failures)
• Verbose filler ("I'd be happy to help!")
• Over-explanation
• Hedging ("It seems like...")
• Thinking bloat

Optimal Prompt Patterns

Pattern 1: Direct Commands

❌ "Could you please help me understand what 2+2 equals?"
✅ "What's 2+2?"

Pattern 2: Act, Don't Ask

❌ "Would you like me to read the file?"
✅ [Reads file, reports result]

Pattern 3: Structured Output

Return as:
- [format specification]

Example:
- [example output]

Pattern 4: Explicit Personality Cues

"Be casual" → More natural responses
"No filler" → Cleaner output
"Just do it" → Action over explanation

Reference Materials

• references/GLM5-OPS.md — Quick reference for prompting patterns
• references/GLM5-DEEP-DIVE.md — Complete model architecture and best practices
• references/GLM5-PERFORMANCE-REPORT.md — Real-world analysis with examples

How to Apply

1. Read the reference materials to understand GLM-5
2. Add the optimization rules to your existing SOUL.md
3. Adjust response lengths and anti-patterns to fit your use case
4. Keep the loop-breaker rule (critical for all GLM-5 agents)
5. Test and iterate based on your specific needs

Validation Checklist

After applying optimizations, verify your system prompt has:

• Loop-breaker rule (3 failures → stop, try different approach)
• Action-reasoning alignment section
• Response length guidelines
• No-filler-phrases list
• Thinking block brevity rule

Based on empirical analysis of 200+ turns of real GLM-5 usage.