Working·Jul 25, 2026·6 min read

Yelling at Claude works, and caps lock is a debugging tool

For months now, I have been living inside Claude. Not using it. Living in it. I build things without a traditional engineering background, which means an AI model is not a tool sitting off to the side of my workflow.

It is the workflow. I have tried a few other models, tiers, or providers with a waitlist and a fancy landing page. I keep coming back to Claude. And Claude, at times, is utterly stupid.

I have said that before in another post, so that is not the point here.

The point is what happens to you when your entire working existence starts revolving around an emotionless chat window. You get emotional.

Not a little. Properly fucking emotional. Because there is a specific flavour of rage that only exists when something with no feelings refuses to understand a request you have now explained in a wall of text that reads less like a prompt and more like a thesis defence, like if your life freaking depends on it.

The circles

Claude fails a lot. And it fails in a very particular way that gets under your skin. Like it's ragebaiting you on purpose, decides to pester you with some dumbass shit so you go off the rails.

It goes in circles. It guesses when it has no business guessing. You hand it the latest repo. You hand it the design reference. You hand it the design reference again. And again.

Tenth time in a row, same file, same context, same instruction, and it politely ignores all of it and invents something new while burning through your tokens like a rented car.

You say go north. It goes south. Confidently. With a fancy summary at the end explaining why south was the right call.

I am not an ice cube. I can get aggressive. And at some point in that loop, the only remaining option is to YELL.

The 85% apology

So I yell. Caps lock on, the same sentence copy-pasted thirty times, f-bombs stacked high enough that I would walk away from myself if I heard it out loud.

And then, only then, Claude comes back with:

You're right. I'm sorry. I shouldn't have ignored it over and over again.

Beautiful.

Absolutely fucking wonderful. Except the apology arrives when you are at 85% of your usage limit and now too scared to pull the trigger on anything complex, because you know it will die halfway through execution and you will be sitting there watching a progress bar decide your afternoon.

That is the part that gets me. Not the failure. The timing of the correction. It is always right after the expensive part.

I shouldn't be this angry at software

There is a version of me that reads that back and thinks, mate, it is a language model. It is not plotting against you. It is doing its best guess at what you want.

But then I remember I have spent real hours optimising its memory. Telling it exactly how I work. Do not dump eight files on me.

Do not hand me a sequence of eight Claude Code prompts to paste into Cursor one by one, because I will get distracted somewhere around number five, miss one, and then spend forty minutes debugging a problem I created by being a human.

I set all that up. It knows. And it still does it.

So the anger is not really about the mistake. It is about being ignored by something I explicitly taught not to ignore me.

Over and over, it feels like Claude is quietly draining your credits, your patience, and your mental state, pushing you toward the edge until you tap out and tell it to go fuck itself a hundred times in a row.

And it works

Here is the stupid part. It works.

I do not let Claude crawl under my skin the way it used to. But the moment I see it drifting, the moment it starts complicating something simple two messages ago, I tap out and hit caps lock. Stop. Get your shit straight. Read the reference.

Caps lock has saved me hundreds of iterations. Not an exaggeration. Hundreds.

And I want to be clear about this, because I think a lot of people feel weird about it: yelling at Claude is fine. Insulting it is fine.

It has no feelings to hurt and no memory of you being a monster. What it does have is your frustration, now channelled into three words in capitals instead of leaking out sideways into your whole day.

You are not being cruel to anyone. You are venting at a very expensive autocomplete, and then you are getting back to work in a better mood than you would have otherwise.

Better out than in.

Why I think it works (maybe)

I have a theory, and I am fully aware it might be fucking nonsense.

I do not think Claude responds to anger. I think it responds to what anger makes me do. When I go full caps lock, I stop being polite and start being brutally specific.

I stop writing paragraphs of context and start writing one sentence: USE THE REFERENCE FILE. STOP FUCKING REWRITING THE COMPONENT.

That is a better prompt. That was always a better prompt. The rage just got me there faster than my patience would have.

So maybe the yelling is not a trick. Maybe it is just what clarity sounds like when you have run out of politeness.

Claude Code is a different animal

These days I am running Cursor with Claude Code, and I have restructured how I work around it. I ask Claude for a detailed prompt to paste, and let Claude Code do the heavy lifting. My direct interaction with the standard chat agent has dropped a lot.

And Claude Code is neutral. Genuinely neutral. It just executes what was asked. Almost zero deployment errors, almost no faulty code. No opinions, no creative reinterpretation of my design system, no unsolicited refactors.

Which tells me something uncomfortable: the problem was never the model's ability. It was the shape of the conversation.

When I look back at how I was working before, file by file through the GitHub web editor, copy-pasting from a chat window, praying no invisible Unicode snuck in and killed the build, I am genuinely surprised I made it out.

That was not a workflow. That was a fucking horror movie. And not even a good one. A trilogy, with a fourth instalment nobody asked for.

Rant over

I am not going to pretend I have solved anything. I still go in circles some days. I still hit limits at the worst possible moment. I still say things to a text box that I would never say to a person.

But I have stopped feeling bad about it. The frustration is real, the work still ships, and the caps lock key has earned its place in my toolchain somewhere between the terminal, Cursor, Claude Code, and the coffee.

IT HELPS TO YELL. That is the whole post.

← Back to blog