Claude, change the “Add to Cart” button to blue
TL;DR
Opusfived is a short interactive game that parodies an AI coding agent asked to turn one "Add to Cart" button blue. The fake Claude's misfires include turning other buttons blue, lecturing about design tokens, spawning subagents and adding a gradient nobody asked for. It even invents a color called "Cyanide Blue" with its own terms of service. The player can only choose between frustrated replies.
The article
Interactive page; only its opening prompt could be fetched. The tldr describes the playthrough from what commenters quote, and the digest covers the discussion.
The discussion
A big, split thread. Half the commenters laughed in recognition and said it gave them physical stress. The other half said current models never act this way and blamed the game's canned prompts. From there the argument spread to over-eager agents, AI coding as gambling, and Claude's house style of writing.
Satire or slander
Many heavy users said the game describes models from a year or more ago, or something you only get by sabotaging the context. Recent Opus or Codex just change the button. Others named Opus 5 as the worst offender, saying one model can one-shot thousands of lines yet burn huge token counts on tiny tasks. Several split the difference, saying the writing style is barely exaggerated even if the behavior is.
It does what I ask it to do, and when it doesn't get it right, it generally turns out there's a good reason for it, which is any of the 100 reasons a human doesn't always get code fixes right on the first try either.
had it one shot the generation of 1000 - 3000 lines of code and tests that work pretty much perfectly first time but then had it chew through six figures of tokens and achieve absolutely nothing useful at all on the seemingly simplest of tasks.
The prompts are the problem
A large group said the game forces you to scold the agent or ask it why, and neither tells it what to do next. Their recipe is to revert, start a clean session and give a scoped instruction, since a bad change left in context keeps steering later answers. One commenter noted the fake Claude had already explained that the button's color token was shared by twenty-odd components, which the player's options then ignore. Pushback came from people who prompt exactly this way and say any human would understand "make the button blue".
Neither of these is an instruction to fix the problem, they're treating the AI like a person and telling it what it did wrong, expecting the implied admonishment to be enough to steer it back. But without an actual instruction, it just goes and does whatever it thinks will help, which is often arbitrary.
Any human with a lick of digital common sense would be able to understand what I obviously mean. How am I supposed to anticipate every single way the AI could possibly misinterpret a simple fucking instruction like that?
Over-helpful agents
The top thread said the real failure today is overdoing it. Models re-run tests, add migration scripts to two-hour-old projects with no users, cast an int to int, and flag unrelated issues. People cope with a project file saying the code is single-user and greenfield, and one reply blamed post-training against hallucination for the endless double-checking. Others reported the opposite problem: Codex is too timid, and Claude sometimes plans a large feature into phases full of churn.
I'll literally add feature A, merge it, then add feature B and it is like "oh no, we'll have to fix up and migrate all of the users using feature A".
Half the time the "one more thing worth noting, unrelated..." isn't even an actual issue, it just brought it up to fish more usage out of me.
Is it gambling?
A long sub-thread argued over the claim that variable rewards keep people hooked on AI tools. One side pointed to near misses, paid tokens and the pull of one more try, and reported all-nighters that produced little. The other side called prompt iteration directed work, not pulling the same lever, with a day's work done in an hour. Someone suggested the vendors should post odds like loot boxes.
And in fact the part that makes it addictive is the near hits where the LLM comes very close to giving you what you want, but not quite there. That keeps you coming back.
You are not executing the same action, but incrementally nudging it in the right direction.
Claudese
Even people who doubted the behavior agreed the game nails how Claude writes, with stock phrases like "worth naming", "load-bearing" and "the honest part". One reader listed every caveat and pushback phrase from the game's spinner. Others said the verbosity wears them down day to day. One found that even writing defensive replies to the fake Claude was exhausting.
I think the key features of claude diction is it sounds like a junior engineer over their skies who is trying to make up for that with extra verbiage mixed with extremely grating SaaS marketing-ese.
The defensive writing is honestly just as exhausting as reading Claude’s output.
Consensus
The voice is accurate. Claude's padded, caveat-heavy prose and its habit of doing more than asked are real and annoying, whatever people think of the button scenario itself.
Dissent
A loud group of daily users called the game outdated or unfair. They said modern models handle a button recolor easily. The bad results come from vague, emotional prompts and messy codebases, not from the model.