I built an AI Resume Builder agent. It takes a name or GitHub profile and generates a complete resume as a Word document. The whole thing runs on claude_agent_sdk. It was a fun experiment.
The SDK gives you powerful built-in tools out of the box. Your agent can read files, run commands, search codebases, and even browse the web. I used these tools:
- Read/Write/Edit - file operations
- Bash - run terminal commands
- Glob/Grep - find and search files
- WebSearch/WebFetch - search and fetch web content
- AskUserQuestion - get user input with choices
The main harness is the Bash tool. It uses Linux to do all the heavy work. Creating documents, formatting, everything happens through terminal commands. It feels crazy when you see it working.
I also used Skills to organize agent behavior. Skills let you define specific workflows. The agent knows when to search GitHub, when to format data, when to generate the final document.
Then I built a clean UI using Cursor. Added live editing with live preview. You can see changes in real time. When the resume looks good, you download the docs file directly. Simple but powerful.
The Cost Problem
First I used Anthropic Sonnet model. It worked great but expensive. For just 3 searches it charged more than 1 dollar. I tried to optimize but still too costly for experimenting.
Then I had an idea. I switched to z.ai GLM4.6 model. It is Anthropic compatible. Same API format works. The cost dropped significantly. Performance reduced a bit but acceptable for experiments. This taught me to think about alternatives.
What I Learned
The claude_agent_sdk is powerful. It handles tools, hooks, subagents, MCP, permissions, sessions. Everything you need to build real agents. After this project I feel confident. I can build any agent using this SDK.
The built-in tools save so much time. No need to write file handling or web scraping from scratch. Just use what is there. Focus on the agent logic instead.
Sometimes the best way to learn a tool is to build something real with it. Theory only takes you so far.
Still experimenting. Not ready to share yet. But the learning was worth it.