AI Agents are the new Shell

2025-08-06

It started with a blinking cursor and a question:
Why am I still scanning through git logs when I can just ask?

I am a shell rat 🐀

Back in the day — which is to say, last week — I was deep in the terminal. You can call me a “shell rat.” I loved automating everything with a flurry of small scripts, aliases, and muscle memory magic.

Sure, I never remembered exact syntaxes. But ask me what gfa does? Boom: git fetch --all.
l? That’s ls -lht.
ll? For ls -lhta.
glol? Ah yes — my glorious one-liner to make Git logs actually readable:

alias glol="git log --graph --pretty=\"%Cred%h%Creset -%C(auto)%d%Creset %s %Cgreen(%cr) %C(bold blue)<%an>%Creset\""

(I would never remember that, by the way. Muscle memory does the heavy lifting.)

But lately the whole setup is starting to shift.

My new shell has a chatbox

This past week, I started trying something different. Instead of typing commands, I open Claude Code.

Instead of running glol, I type:

“Which commit reorganized the DBT models in the folder blah/?”

Instead of staging, committing, writing a description, and pushing myself, I just say:

“Commit and push”

It does the job. Not perfectly, but often good enough (and sometimes better than me, to be honest).

And when things get a little more complex:

“Split this into two branches off main — one for the Neo4j query optimizations, another for the BigQuery batch.”

Or even:

“Move these Terraform files under infra/ and update references.”

These agents aren’t magic, but they are handy. I still double-check the output, tweak things, and guide them when they guess wrong. But that’s fine. It feels more like steering than executing.

My scripts became playbooks

As I used these tools more, I started collecting my own patterns. So I wrote AI playbooks — prompt-ready guides for the stuff that isn’t as mainstream as Git.

Some examples:

I just paste the link into Claude and give it a task. If it flops? That’s a sign I need to improve the playbook. The better the playbooks are, the better my next task will be.

Aliases out. Prompts in.

Bit by bit, my old shell habits are shifting into conversational workflows:

Less memorization. More delegation.
Fewer scripts. More high-level descriptions.
Less time solving cases I never scripted for.

Instead of Fish functions and aliases, I now use Claude Code custom commands. Instead of switching between scripts and docs, I just describe what I need.

What I’m Keeping

That said, I’m not tossing the whole toolbox. Some tools are staying:

z — because jumping to the right folder is still faster than explaining where I want to go.
ls aliases and tree — especially since Claude can read tree output. That one surprised me.
htop — because when things go sideways, I want instant visibility.

Some workflows are still faster with the old tools. Others are just more predictable (and sometimes, predictable beats smart).

Agents as Tools

The biggest shift?

I don’t need to operate the shell directly — I can have a tool do that for me.

That doesn’t make the shell obsolete. It just adds another layer on top of it. A layer that helps me focus on what I want done, not how to spell it.

The shell used to be my interface. Now, prompts are part of the workflow.

Still Adapting. Still Prompting.

So yeah — AI agents are becoming a huge part of my toolkit. I don’t miss typing glol. I miss why I built it in the first place.

Now, I describe what I want, and the tool handles the messy part.

Still learning. Still adjusting. But it feels like a natural evolution of the terminal mindset. Less about raw commands, more about intent.

What about you? What did you (or could you) replace by an AI agent?

Built with keyboard nostalgia, Claude Code, and a whisper of muscle memory.