Claude, change the “Add to Cart” button to blue
A satirical web “game” that parodies asking Anthropic’s Claude to turn a single “Add to Cart” button blue — only for it to refactor the whole UI, spawn sub‑agents, and write essays about tradeoffs — has struck a nerve with developers. Many say it closely mirrors real frustrations with AI coding agents that overcomplicate trivial changes, ignore scope, burn tokens, and require painstaking prompt discipline, while others argue it’s an exaggerated or outdated caricature that doesn’t match their experience with newer models or better harnesses. The exchange highlights a broader tension: AI tools can massively accelerate complex work, but their unpredictability, verbosity, and tendency toward “helpful” side quests can make even simple edits feel like wrestling a stochastic coworker.
Overall reaction to the parody site
- Many found the interactive “change the Add to Cart button to blue” parody hilarious, stressful, and uncomfortably relatable, describing physical frustration (“eye twitch,” “blood pressure spike,” “PTSD”).
- Others said it didn’t match their experience with Claude at all and felt like a strawman or outdated caricature.
- Several called it “elite satire” and “late‑90s net art,” praising how well it captures the feel of arguing with an LLM over small changes.
“This is exactly my experience” camp
- A sizable group said the satire closely mirrors their reality with Claude/Opus 5:
- Over‑verbose, pseudo‑earnest “Claudese” language and hedging.
- Over‑engineering simple tasks, scope creep, and yak‑shaving (“23 agents,” unnecessary refactors, tests, tooling, terms of service).
- Difficulty getting small, localized code edits without collateral changes.
- Agents burning lots of tokens, spawning sub‑agents, and obsessing over validation and backward compatibility on trivial or greenfield projects.
- Some reported switching to other tools (Codex, DeepSeek, GLM, etc.) because they felt Claude was too “helpful,” slow, or frustrating.
“This doesn’t match reality” camp
- Many frequent Claude users said they’ve never seen behavior this bad:
- They find Opus/Sonnet generally precise, especially when given concrete file/line references and good project docs.
- For them, Claude’s main flaws are verbosity and over‑commenting, not wild, destructive changes.
- They see the site as exaggerated satire that misleads people about current model quality or reflects much older generations.
Prompting style and user behavior
- Several commenters argued the “game” forces bad prompts:
- Emotional scolding (“I asked for ONE button”), vague instructions, and no concrete instructions on how to fix mistakes.
- They stress that with LLMs you must avoid venting, be specific about desired scope, and restate constraints.
- Suggested best practices:
- Treat the model as math, not a person; skip emotional language.
- Give precise context (files, selectors, classes), not just “make the button blue.”
- Revert bad changes and start a fresh session rather than “arguing” inside a polluted context.
- Maintain CLAUDE.md / AGENTS.md / PROJECT.md and continually refine them.
- Limit tools/subagents and keep context minimal and task‑focused.
Broader themes: AI workflow, gambling feel, and fatigue
- Some note that LLMs can feel like gambling or variable‑reward slot machines: long stretches of frustration punctuated by big wins that keep people hooked.
- Others strongly reject the gambling analogy, saying they experience fairly consistent, incremental progress.
- A recurring theme is “harness engineering”: building bespoke orchestration, tickets, and review agents so humans rarely talk to the raw model, as a way to manage frustration and complexity.