How I code with agents, without being 'technical'

The author says they have spent billions of tokens in the terminal, letting an agent write code while they watch the output and learn how projects work. They describe this as their way of learning to program and as part of a new technical class shaped by AI agents.

Reading notes

#

• They use a CLI exclusively and prefer the terminal over web interfaces because it is more capable and lets them see the agent work.
• They start with context, switch to spec mode, question assumptions, link docs and repositories, and then let Opus 4.5 run with high autonomy.
• They learn by building ahead of their ability, then using the gaps and bugs they hit as learning opportunities.
• They keep an agents.md file in a local repos folder to define how new repos should be set up, how to use GitHub, how to commit, and which account to use.
• They want end-to-end tests on everything because they now notice how many bugs could have been caught earlier.
• They use the Droid GitHub app, Telegram, and Slack so they can work on repos from phone or while away from their desk.
• They have learned bash commands, CLI tools, VPS basics, and portable personal CLIs such as Gmail and Linear tools.
• They see AI agents as a new programmable layer of abstraction, similar to how Webflow, Zapier, and Airtable were the abstraction layer in no-code.
• They read other engineers’ workflows, clone and adapt open-source projects, and prefer simpler systems when those already work.
• They frame the process as exploration, where they can try ideas, throw things away, and keep shipping without needing to be an expert programmer.