Building crumpits.com with Agentic Development in OpenCode

This entire site was written by AI agents. Every component, every route, every Prisma query — none of it was typed by hand.

The tool? opencode — an open-source AI coding assistant that runs in your terminal. You describe what you want, and it writes the code. It sounds magical, and sometimes it is. But the first few sessions? Honestly, they were frustrating.

The hard part

When you start with an AI coding agent, the learning curve isn't about the code — it's about the interface. You have to learn:

• What the tools do — opencode can search files, read code, edit files, run commands, and launch sub-agents. Each has a specific strength, and using the wrong one wastes time.
• How to prompt effectively — vague instructions produce vague (or wrong) results. You learn to be specific about file paths, conventions, and constraints.
• When to trust the output — the agent will confidently generate code that references APIs that don't exist. Catching that takes experience.

I spent the first day fighting the tool more than building the site. It felt slower than just writing the code myself.

The breakthrough

Somewhere around the third session, something clicked. I started to internalize the patterns:

• Need a new page? Describe the route, the data source, and which components to reuse. Done in 30 seconds.
• Need a consistent component? Show the agent one existing example, then ask for another. It replicates the style perfectly.
• Need to refactor? Describe the change at a high level. The agent handles the multi-file update.

The key insight is that opencode isn't magic — it's a multiplier. If you know what you want, it removes the typing friction. But you need to know what you want first.

Tools and agents

What made the difference for me was learning about the different agent types. opencode has specialized agents:

• Explore agents — fast, good for searching the codebase and finding patterns
• General agents — slower but more capable, good for implementing features
• Task agents — focused sub-agents that handle specific work independently

Using the right agent for the job makes everything smoother. Explore for research, general for implementation, task for parallel work.

The verdict

The experience curve is real: hard at first, easier with practice, and genuinely productive once you're past the hump. If you're trying agentic development for the first time, expect friction. Push through it. Read the docs. Learn what each tool does. The payoff is worth it.

Try opencode yourself. It's free, open-source, and works with any model you want.