How the Work Actually Happens

The most effective workflow for AI-assisted coding starts with a CLI (command-line interface). Terminal over web interfaces, always. It’s just more capable as a general agent, and you get to watch it work.

Here’s how a typical session looks:

It starts with an idea, a pain point, or an issue that feels like it could be solved with code (basically everything these days). So you spin up a new project in your CLI agent.

The first step is usually just talking to the model a couple of times to start feeding in context about what you’re trying to do. Then switch into spec mode (or plan mode) to start getting a plan together for what you want to build.

In spec mode, question everything. “What is this?” “Why would you use that over this?” “Can’t you do it this way instead?”

Link docs and GitHub repos for the agent to explore.

Then let a powerful model like Claude Opus with high autonomy just rip. Watch the stream, see what’s happening, and jump in when there are errors. You might question a decision or guide it down a different path.

Start the server, test it, give feedback, and iterate.

The key pattern here is: build ahead of yourself first. Just try to build the thing. All of the gaps and issues that come up are opportunities to learn. Is that something common across other repos that should become a template? Should it go into an agents.md file that follows you around and does the same thing on every project?