LLMs are dumb - albeit vast and surprisingly accurate - next-word predictors.
On their own, they're interesting, but necessarily very impressive or useful.
They need a harness, and skills to be useful - which is to say that you take their base 50% accuracy and run it in a loop with hand-picked context, and let them try a few times until they get it right.
Oh, and you give them a special machine-parsable tool vocabulary that the harness can run and then provide feedback.
Not that there's one right way, but here's an easy path:
-
Install Ollama: https://ollama.com/download
-
Get Ollama Cloud: https://ollama.com/pricing
- Gives you access to GLM, Kimi-K, Minimax, etc (like Opus / Sonnet / Haiku)
- Free is enough to try, $20/month will get you far
-
Sign in
ollama signin
-
Install Node, OpenCode & Claude CLI
curl -sS https://webi.sh/node@lts | sh source ~/.config/envman/PATH.env cd /tmp npm install --location=global opencode-ai npm install --location=global @anthropic-ai/claude-code
-
Launch GLM-5 and ask it to install these skills
ollama launch # down arrow to OpenCode # right arrow to select models # hit enter on glm-5
Clone https://github.com/therootcompany/skills to ~/Agents/skills and use agent-init touse the walk me through agent initialization.