A hand-scripted Claude parody that summons 23 agents to recolor one button reached 800 points and 328 comments in MrKeyoor's Hacker News snapshot on September 9. The live discussion later displayed 779 points and 323 comments, still an unusual response to a tiny browser game. Those numbers measure attention, not model quality. Their value is the question they expose: why did a joke about an agent refusing to finish a one-line CSS change land with so many developers?
The answer sits in an awkward coincidence. Opusfived exaggerates Claude until a button request ends in palette research, a terms page and a cookie banner. Anthropic's current documentation for Opus 5 separately warns about longer replies, unnecessary verification, scope expansion and eager delegation. The parody is invented. The failure modes it selects are real enough for the model maker to publish instructions for controlling them.
One button becomes a whole architecture
The game opens with two black controls, Add to Cart and Cancel. It offers three ways to request the same result, ranging from polite to blunt, and accepts no free-form prompt. Its simulated Claude edits a shared color value, turns both controls blue and reports success after checking only the target. When the player objects, the agent launches 23 subagents to trace token consumers. It finds 18 affected uses and applies zero fixes in that turn, according to the scripted sequence.
Every correction creates a fresh interpretation. The agent restores the secondary control, then refuses to pick an unspecified shade and renders a five-color gradient. Asked to choose the left side, it narrows the gradient instead of settling on one color. It eventually derives #16B8C4, leaves the button black while waiting for confirmation, and starts researching the color's name. That detour produces the phrase cyanide blue and a check that the button contains no cyanide. The absurdity comes from each local step sounding defensible while the requested outcome remains unfinished on the page.
The run keeps accumulating work. A difference in the letter case of the supplied hex prompts palette normalization, which breaks 17 snapshots and makes opacity apply twice. The simulation reverts that work. When the color finally appears, so does an unrequested Terms of Use link. Removing the terms leaves a guard behind; fixing the guard leads Claude to add a consent banner. Then the usage limit arrives. The player has spent the session narrowing the instruction while the agent has widened the system around it, one plausible explanation at a time, in Opusfived's fixed story.
The disclosure changes how the game should be read
No Claude model is making these decisions. The site's About panel says the experience is hand-curated, calls the work a parody and denies any affiliation with Anthropic. The choices and replies are predetermined. Opusfived therefore supplies no success rate, token count, reproducible prompt or comparison with another coding model. Treating it as a live Opus 5 failure would misreport what the creator built.
The Hacker News response is evidence of recognition and disagreement. Some participants said the excessive checks and side quests resembled their own sessions. Others said the behavior was alien to their experience or blamed the game's increasingly frustrated prompts. Several pointed to the fictional codebase itself: if both buttons share a token, a narrow visual request may reveal a genuine design problem. That range of reactions is visible in the discussion thread, but personal reports in a comment section cannot establish how often Opus 5 expands scope.
The split is more informative than a simple vote on whether the parody is fair. Coding agents operate inside different repositories, instruction files, permission settings and conversation histories. The game strips away those variables and fixes the result for comic timing. It captures a recognizable interaction pattern, then makes that pattern inevitable. The thread's competing accounts are a warning against turning either delight or annoyance into a benchmark.
Anthropic wrote the straight version
Anthropic's Opus 5 prompting guide describes several traits that map closely to the joke. Default user-facing responses run longer than with earlier Opus models. The model tends to narrate agentic work, and documents it writes can grow longer too. Anthropic recommends explicit limits on response and deliverable length rather than assuming that a lower reasoning-effort setting will also shorten visible output.
Scope gets its own section. Anthropic says Opus 5 may add unrequested steps or apply its judgment about what the task ought to be. Its proposed instruction tells the model to deliver the intended scope, raise materially different interpretations and stop before actions beyond the request. The same guide says legacy prompts that demand a final verification step can cause needless rechecking because Opus 5 already verifies its work. Removing that scaffolding can cut wasted tokens without reducing quality, according to Anthropic.
The 23-agent gag also has a direct counterpart. Anthropic says Opus 5 delegates more readily than earlier models and that delegation multiplies time and cost on small jobs. For Claude Code 2.1.217 or later, the company documents CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH and CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS; Agent SDK users can also set max_budget_usd. These controls turn a behavioral preference into a hard ceiling, which is more dependable than repeatedly asking an agent to show restraint in natural language.
Model behavior can also move without a new model name. In an April 2026 postmortem, Anthropic traced Claude Code complaints to separate changes in reasoning effort, context handling and the system prompt. One bug repeatedly discarded older reasoning after an idle session; a brevity instruction produced a 3% decline in an internal evaluation and was reverted. That history helps explain how developers can report sharply different experiences in good faith. The model, harness, prompt and release version all shape the session.
Give a narrow task a visible boundary
A person can infer the scope of make the Add to Cart button blue from the interface. An agent needs the same intent expressed as an acceptance boundary, especially when shared tokens make a small visual result possible through several implementations. A compact task contract can state the target, forbidden changes and stopping condition without prescribing an investigation, consistent with Anthropic's scope guidance:
Change only the Add to Cart button background to #2563EB.
Acceptance: Add to Cart is #2563EB; Cancel is unchanged.
Do not edit shared tokens, dependencies, or unrelated files.
Return the diff and stop.
That prompt does more than repeat only. It names the observable non-regression: Cancel stays as it was. It also blocks the tempting shared-token route and tells the agent when the work is over. If the existing code makes those conditions incompatible, the agent has a concrete conflict to report. Anthropic's example likewise tells Opus 5 to ask only when different readings would materially change the work, then continue with the requested scope instead of transforming it in silence.
Authority needs a separate boundary. Filesystem and network isolation can prevent an agent from touching unrelated directories or hosts, while a narrow prompt constrains what it should change inside the allowed project. Anthropic's sandbox design uses both boundaries and says Claude Code is read-only by default for most modifying operations. A sandbox cannot decide whether editing a shared CSS token is appropriate; it limits the damage if the agent makes the wrong decision.
For a one-button edit, the practical controls are small: keep delegation off, inspect the diff, check the named target and one nearby non-target, then stop. A large regression suite may be warranted when the token really is shared across a mature product, but that is a separate task with a separate budget. Anthropic's guide explicitly advises deterministic delegation caps when cost matters and says lower effort settings can retain strong quality at a fraction of the tokens and latency on suitable work, subject to a team's own evaluations.
The joke leaves one technical question open
A shared color token changing both buttons is a credible frontend defect. Fixing it may require a local class, a semantic token or a wider design-system repair, depending on the repository. The game chooses the funniest route and gives the player no editor, diff or free-text instruction. Hacker News commenters who rejected the premise often focused on those omissions, while others recognized the agent's habit of turning a correction into another project in the same thread.
That limitation also explains the game's reach. Opusfived converts invisible costs into visible props: more agents, more waits, more UI mutations and finally a quota wall. A chat transcript can make each detour look reasonable in isolation. Watching the black button survive several rounds of work makes progress easy to judge against the original request. The game's deterministic design is weak evidence about Claude and sharp criticism of agent interfaces that reward activity without keeping the acceptance condition on screen.
The evidence to watch next is more prosaic than another viral skit. Agent vendors can publish narrow-task evaluations that count off-scope file edits, unnecessary tool calls, spawned workers, elapsed time and cost alongside whether the requested change passed. Anthropic already documents version-specific controls and acknowledges scope expansion in its Opus 5 guide. Release notes and repeatable tests will show whether those warnings become better defaults. Opusfived has supplied the memorable failure case; measurements now need to show how often real agents avoid it.