Claude, change the “Add to Cart” button to blue

Who Should Read

Developers frustrated when coding agents change more than they were asked to

Core Mechanics

• A game about turning one button blue
  • The task is to make the "Add to Cart" button blue.
  • The constraint is not to let Claude change anything else.
• The source text is just a short blurb
  • It contains only those two sentences, a 'Let's go' button, and an 'About & Terms' link.
  • The in-game dialogue and rules are not in the source.

Community Reactions

• Satire that is not far from reality, per one commenter
  • They describe getting 1,000-3,000 lines of working code and tests in one shot.
  • The same person also burned six figures of tokens on the simplest tasks with nothing useful to show.
• Opinion that models over-verify
  • One comment says models write a Rust variant just to prove a working solution.
  • Another had to stop a model from benchmarking latency across 50 runs on a first version.
• Pushback that it does not match their experience
  • Several commenters say they have never seen Claude behave this way.
  • Some suspect the example is contrived.
• Experience that Codex can trace its reasoning
  • One commenter says asking "why did you do this" gets the decision traced back.
  • They usually found the cause was a wrong spec or a skill file and fixed the instructions.
• Opinion that doing it by hand is faster

How to Apply

• Add a scope limit to your instructions
  • When asking a coding agent, state "only this element in this file; touch nothing else".
  • Run git diff afterward to check for unintended changes.
• Make asking "why" a habit
  • When an odd change appears, ask "why did you do this" to see the reasoning.
  • If the cause is a skill file or instruction, fix that document.

Try It

# Paste into Claude Code or ChatGPT (scope-limiting prompt, a simple version of the site's scenario)

Do only this: change the color of "{{element name}}" in {{file path}} to {{color}}.
Rules: no other elements, no other files, no refactoring, no added tests, no new color definitions.

When done, report only the changed files and line counts.

Next to try: run git diff and confirm only the intended change is there.

Terminology

Related Papers

• Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

  Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.

• Harness Engineering 101: The LLM is the brain, the harness is the body

  An agent is a loop managing a JSON array; 16 chapters build one from scratch

• Harness2 by Google Research – Recursive Agent Harnessing

  Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.

• Ousted Savannah police officer gave ICE agent husband access to Flock system

  A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired

• Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling

  Running Claude Code agents in tmux works well, but limits appear around four agents.

Related Resources

• https://opusfived.dev/