💻 Claude Code: Agent in the Terminal

Claude Code: Agent in the Terminal: Claude Code is like handing a very capable new team member a keyboard and telling them "you may touch these files, run these commands, but ask

Claude Code is like handing a very capable new team member a keyboard and telling them "you may touch these files, run these commands, but ask before doing THAT" — the mechanism is exact: permission modes (ask-every-time vs. auto-approve for read-only vs. full access) are literally the boundary of what the new hire is allowed to do without a manager present, the same distinction between a junior who needs sign-off on every commit and a senior with merge rights. Here is the question worth sitting with: if Claude Code can read a failing test, run it, and edit the fix itself, why would you ever want it to ASK before running a command instead of just doing it? Because the blast radius of an autonomous agent scales with what it's allowed to touch — a read-only permission mode is safe by construction (nothing to undo), while a mode that can run arbitrary shell commands or push to a branch needs the same review discipline you'd apply to a new hire's first unsupervised week. Java comparison: this is access control the same way a role-based permission system draws a line between "can read this resource" and "can write/execute" it — you don't grant the wider permission just because the narrower one works fine. The QA stake: an agent given unreviewed write+execute access to a shared CI runner or a production-adjacent branch, without a human checking its diffs, is a supply-chain-shaped risk — the same category of concern covered on this platform's own security page, just with an AI as the actor instead of a human attacker.

The Read → Run → Fix → Re-run Loop

Once given permission, Claude Code reads the failing test file, executes it, reads the actual error output, proposes an edit, and re-runs — the same debugging loop a developer does manually, but the agent does it in the same terminal session without you retyping every command in between.

Meta-example — this very platform: this project's own repository has a CLAUDE.md file at its root, a written constitution the agent reads at the start of every session so it doesn't have to be told the same project rules over and over. A representative excerpt (not the real file, a stand-in of the same shape):

How Does CLAUDE.md Actually Shape the Agent's Behavior?

When a session STARTS…

When a session STARTS, Claude Code AUTOMATICALLY reads the CLAUDE.md file at the project root — this means you don't REWRITE the same rules in EVERY prompt.

The "content changes stay within src/data/*.js" rule…

A rule like "content changes stay within src/data/*.js" PREVENTS the agent from ACCIDENTALLY touching component/config files — the boundary is defined in the FILE, not the prompt.

The "every code block pairs with a practice/animation" rule…

The "every code block pairs with a practice/animation" rule makes the agent RE-CHECK its OWN output against this rule even when it THINKS a task is DONE.

The "run the build after every change" rule…

The "run the build after every change" rule FORCES a verification STEP that CLOSES the gap between "I made the change" and "the change ACTUALLY works".

This file follows the SAME logic as a CONTRIBUTING.md…