Free GLM5 Paired with Claude/Codex Skill Building Guide

GLM-5 is Zhipu's latest MoE model, with a total of 744B parameters (40B active). It's a significant upgrade from GLM-4.5's 355B, with pre-training data increased to 28.5T tokens. It's specifically designed for complex systems engineering and long-term agent tasks, excelling in coding and Agentic workflows.

Ollama Cloud now supports glm-5:cloud, with a context window of 198K, fast speed, and a free starting point (capacity has been optimized; there was a brief rate limit early on, but it has now been increased).

Now I'll teach you how to get GLM-5 + skill building for free.

Step 1: Install Ollama and Pull GLM-5

Download the latest Ollama (https://ollama.com/download), ensuring the version supports cloud models.

Log in to your Ollama account (required to use cloud models).

Pull the model:

ollama pull glm-5:cloud or directly

ollama run glm-5:cloud The first run will prompt you to authorize cloud usage; agree to it. After successful completion, you can directly converse with GLM-5 in the terminal.

The free version has rate limits, suitable for testing/daily use; upgrade to a paid plan for high-frequency use.

Step 2: One-Click Bridge to Claude/Codex Skill Environment

Ollama launch allows GLM-5 to seamlessly become the backend brain for these tools:

Claude Code

ollama launch claude --model glm-5:cloud or manual configuration

export ANTHROPIC_AUTH_TOKEN="ollama"export ANTHROPIC_BASE_URL="http://localhost:11434"export ANTHROPIC_API_KEY="" # Must be Empty Then start

claude --model glm-5:cloud If successful, you will see a welcome screen similar to the following:

✦ Claude Code v2.1.xModel: glm-5:cloudContext: 198KWelcome back!

Codex

ollama launch codex --model glm-5:cloud

OpenClaw

ollama launch openclaw --model glm-5:cloud

Step 3: Build Skills Efficiently in 5 Steps

A skill is a folder, and the core file is SKILL.md, which uses Markdown format + YAML frontmatter metadata. Other optional directories:

• scripts/: Stores executable scripts (Python, Bash, etc.).

• references/: Reference documents, template files, which Claude can load as needed.

• assets/: Resources such as icons and fonts.

Design Principles (Key to Efficiency):

• Progressive Disclosure: YAML only contains core descriptions (allowing Claude to quickly determine if it's applicable), the main content is in the MD body, and additional files are loaded as needed. Avoid context explosion.

• Composability: Skills should be compatible with other skills and should not assume exclusivity.

• Portability: A skill can be used universally in Claude Web, Claude Code, and API.

• MCP Integration (Optional): If there is an MCP (Managed Claude Platform) server, the tool calling of skills can be enhanced (such as API access).

Efficient Tips: Start with simple use cases. Choose 2-3 programming tasks that you repeat daily (such as "generate REST API templates" or "debug Python errors"), and avoid building complex skills from the beginning.

Use Claude Code itself to assist in building (meta-skill technique: let Claude help you generate the skill framework). Run claude --model claude-4-opus (or GLM5) in the terminal, and then build it conversationally.

1. Plan Use Cases and Structure

• Define Goals: For example, "Create a skill to help generate efficient Python Flask API architecture, including routes, error handling, and test templates."

• Brainstorm with Claude Code: Enter /plan or directly say "Help me plan a Flask API generation skill, including YAML and MD structure." Claude will output a draft.

• Efficient Tips: Specify "Keep it concise, YAML < 200 words, MD < 1000 words." Record key decisions, such as input parameters (e.g., API endpoint list) and output format (JSON or code files).

2. Create Skill Folder

• Create a new folder under ~/.claude/skills/ (default path), such as flask-api-generator.

• Create SKILL.md and add YAML frontmatter:

---name: Flask API Generatordescription: Generates complete Flask API structures with routes, error handling, and tests.version: 1.0author: Your Nametriggers: ["flask api", "generate rest api", "python web app"]dependencies: ["flask", "pytest"] # Optional, Claude will check the environment---

• Efficient Tips: Use trigger words to allow Claude to automatically discover skills. Avoid generalized descriptions and focus on specific scenarios.

3. Write Core Instructions (MD Body)

• Write detailed steps after YAML:

`## Step 1. Ask the user for API specifications: endpoints, methods, parameters.2. Generate app.py: Contains Flask app, routes, and basic error handling.3. Generate tests.py: Unit tests using pytest.4. Output file: Save to the current directory using the /write_file command.

Example Input- User: Generate a user management API, including GET /users and POST /users.

Best Practices- Always use type hints and docstrings.- Prioritize security: Add input validation.`

• Add script examples: Place generate_routes.py under scripts/, which Claude can call.

• Efficient Tips: Use numbered steps + examples. During testing, let Claude "simulate running": say "Use this skill to process a sample task" and iterate to optimize.

4. Add Optional Components

• Scripts: Such as Python scripts to automate part of the logic (e.g., use Jinja2 to generate templates). Claude can execute it through /run_script.

• Reference Files: Place references/best_practices.md, describing Flask security specifications.

• Assets: Such as template files assets/base_app.py.jinja.

• Efficient Tips: Only add necessary items. 80% of skills only need SKILL.md. MCP users can integrate external tools (such as database connections).

5. Test and Iterate

• Load Skill: Enter /load_skill flask-api-generator in Claude Code or let it automatically discover it.

• Test: Enter a trigger task and observe the output. Check context usage (use /context to see if it is overloaded).

• Iterate: If it fails, use the "Heal Skill" idea (community technique): analyze the error and update MD (such as adding a "handle missing dependencies" step).

• Efficient Tips: Use multi-turn conversations to test. Record logs: /log on captures interactions, making it easier to debug.