Part 1: From ChatGPT to opencode — Doing Work, Not Just Asking

If you already use ChatGPT to help with code, you know the dance: paste a question or a snippet, read the answer, copy it back into your project, hope it fits. It works, up to a point.

opencode works differently, and once that clicks, it’s hard to go back. This is the first post of a short series introducing you to AI-assisted coding (no programming background needed). It’s a big-picture look at what changes when you stop posting questions to an assistant and start handing work to an agent. In Part 2 you’ll install opencode so you can follow along with the rest.

The difference in one sentence

ChatGPT is a brilliant co-worker who only talks. opencode is that same co-worker, handed the keys to your project. It can open your files, change them, run commands, test its own work, and come back when it’s finished.

Practically, that changes your role. With ChatGPT, you are the one doing all the choreography: pasting code, saving edits, running the tests, checking if it worked. With opencode, the agent does the choreography while you direct it and check the result.

What an agent can do that a chat can’t

Here’s the concrete gap for “do work” tasks:

• It edits your actual files. You don’t paste a diff in; tell it what to change and it changes the real files on disk.
• It sees your whole project, not just one snippet. It reads surrounding code, so edits fit how your code actually is, not how its training data roughly pretends.
• It runs commands. It can run a script, run a test, or check something and read the output.
• It can go back and forth on its own. It notices a step failed, fixes it, and retries; you don’t relay every message.

When ChatGPT is still the right tool

An agent is not always better. Keep ChatGPT (or a quick web chat) for these:

• Quick lookups: “what does this error mean?”, “remind me of the syntax for X.”
• One-off questions: you want an explanation, not a change.
• Things you don’t want touching your files: brainstorming, discussing approaches, drafting ideas before anything is committed.

Positive rule of thumb: if the outcome you want is a change to your project, reach for an agent. If it’s just knowledge, a chat is fine.

How to ask an agent for work (the prompting shift)

Chat prompts and agent prompts sound different. Compare:

• Chat-style: “Here’s my query, is it correct?” You want a judgment back.
• Agent-style: “My query is slow and I have this data. Make it faster, and tell me what you changed.” You want a done thing.

A few patterns that make agents perform well:

• Give it a goal, not a request for an opinion. State the outcome: “clean this CSV and give me the top 5 customers by region.” Leave it to work out how.
• Let it use the project. If opencode is running from your project folder, it can read what’s relevant. Run /init in a project once so it builds an understanding first.
• Say what “done” looks like. “Fix the X, and run the test to confirm.” It knows when to stop.
• Check its work. Read the diff or the summary it gives. You don’t have to understand every line, but skim what it changed before trusting it.
• If it stalls, add constraints. “Don’t touch this file”, “use Excel formulas, not Python now”, “keep it simple.”

Your role hasn’t vanished: it changed

The valuable part of your job isn’t typing. It’s deciding what needs to happen and judging whether the result is right. Those two things are exactly what you keep. OpenCode handles the typing and the fiddly mechanics.

In Part 3, we’ll walk through one whole real task end to end (a messy CSV your inner data person will recognize), so you can see the pattern in action rather than abstractly.