glm-vision

For use only with Pi! get-your-pi-here

get it: npm install -g @mariozechner/pi-coding-agent

A skill that gives text-only GLM-4.7 vision capabilities by delegating to GLM-4.6v. A funny thing about GLM-4.7 is that it is CONVINCED it's name is claude, so you'll notice that claude is specified in the tool description even though this is targeted only at glm-4.7 because of glm-4.7's identity crisis.

What it does

Text-only models can't process images or videos. This skill provides a simple pattern: call pi with GLM-4.6v and a reference prompt tailored to the task.

pi @screenshot.png --provider zai --model glm-4.6v -p "<prompt>"

That's it. The skill is just a lookup table mapping content types to pre-written prompts.

Structure

• SKILL.md - Minimal instructions + task→reference lookup table
• references/ - 25 prompt templates covering all vision use cases

The reference prompts are organized by task type:

• UI-to-code: Generate code, specs, or prompts from screenshots
• OCR: Extract text, code, or multi-language content
• Error diagnosis: Analyze error screenshots with optional context
• Diagrams: Architecture, flowchart, UML, ER, general technical
• Charts: General analysis, trends, anomalies, dashboards
• Comparison: UI diff between two screenshots
• Video: General analysis, tutorial step extraction

About the skill description

The description in SKILL.md is intentionally emphatic:

CLAUDE and GLM-4.7 (text-only) CANNOT process images or videos - you MUST ALWAYS delegate vision tasks to GLM-4.6v via this skill...

This wording was written by GLM-4.7 itself through iterative refinement until the model correctly triggered the skill 100% of the time on first try in testing.

GLM-4.7 is convinced it's Claude as the model name). This is relevant because the description explicitly addresses "CLAUDE" - the model the agent believes itself to be. The aggressive, self-referential phrasing ("this is your ONLY way to understand visual content", "NO EXCEPTIONS") works because the model wrote instructions it would actually follow.

Why this approach

This skill replaced a dedicated MCP vision server. The original server had tools for each use case (OCR, UI-to-code, diagram analysis, etc.). This skill achieves the same coverage with:

1. One command - pi @file --provider zai --model glm-4.6v -p "..."
2. Reference prompts - Pre-written prompts for each task, loaded only when needed
3. Minimal SKILL.md - Just a lookup table, keeps context lean

The agent reads SKILL.md, picks the right reference file from the table, and constructs the pi command. No server, no dependencies beyond pi.

Requirements

• pi installed and configured
• z.ai API key set up in pi (see pi docs for provider configuration)

Usage

Install as a skill in your agent's skills directory, or reference directly. The agent will automatically use it when visual content is mentioned.

License

MIT