Open Source Project "done": Satirizing AI Tool Metric Obsession With a Single Word

An Open Source Project That Takes Satire to the Extreme

In the world of AI coding tools, a new plugin seems to appear every few weeks, each claiming to save tokens, reduce context pollution, or optimize your workflow. These tools have proliferated to the point of becoming a predictable formula for attention. A recent open source project that went viral on Reddit pushes back against this "cost-reduction anxiety" in the most restrained yet absurd way possible.

The project is called done — a Claude Code plugin. Claude Code is Anthropic's AI coding assistant for developers, which supports a plugin ecosystem for extending functionality. This ecosystem has spawned a wave of third-party tools, especially those focused on token optimization. Since large language models like Claude charge per token, complex agent tasks can easily consume tens or even hundreds of thousands of tokens. That cost pressure has given rise to a flood of plugins claiming to "trim context" or "compress prompts" — exactly the culture that done is mocking.

done does exactly one thing: respond to any request with a single word — Done. No code generation, no tool calls, no reasoning or analysis. Just a clean, confident declaration that the job is finished.

The author's tongue-in-cheek announcement: "You can all stop now. I've found the ultimate solution."

A "Breathtaking" Performance Benchmark

The author deadpans a set of benchmark metrics, pushing the performance art to its peak:

• Output tokens: Reduced by 99.9%+ under any workload
• Tool calls: 0 (down from multiple)
• Response time: O(1) — constant regardless of task size
• Context pollution: None
• Actual work completed: None

It's worth noting the author's use of Big O Notation in the response time row. O(1) denotes constant time complexity — meaning the operation takes the same amount of time regardless of input size, and is the theoretical optimum in algorithm analysis. In AI inference, token generation time typically scales linearly with output length (O(n)). Since done always replies with a single word, it genuinely achieves the theoretical ceiling on this metric. That's the brilliance of the dark humor: using perfectly correct technical terminology to seriously describe a completely useless system.

The punchline is that every metric looks perfect — except the last line, "Actual work completed: None," which exposes the meaninglessness of all the others. This is the sharpest jab at a certain class of "optimization tools": when we chase surface metrics like token counts and response speed, it's easy to forget that the entire point of a tool is to actually get something done.

The Obsession With "Authenticity"

The most delightful detail in the project is the author's insistence on the "authenticity" of each response:

"Before you ask: yes, every Done. is genuinely generated end-to-end by a frontier model. v2 had this string hardcoded, but beta testers said it 'didn't feel authentic,' so I fixed it."

This cranks the satire up to maximum. "Powered end-to-end by a frontier model" has become a standard marketing phrase since 2023 — no matter how mundane the feature, slapping "GPT-4 powered" or "Claude-native" on it provides instant prestige. done deconstructs this perfectly: it burns real compute from a world-class language model to generate a string that any if-else statement could produce in microseconds, perfectly exposing the absurd gap between "AI-native" rhetoric and actual value. Making a frontier model expend compute to generate the single word "Done." is itself the ultimate joke about wasted resources.

The Installation Instructions: Serious Enough to Be Funny

The author even provides a complete, properly formatted installation guide, indistinguishable from a real Claude Code plugin:

/plugin marketplace add handgemacht-ai/done
/plugin install done@done
/config  // set output style to done:done

The repository lives at github.com/handgemacht-ai/done. This "playing it completely straight" seriousness is the core of the project's humor — packaging a function that does absolutely nothing in the full regalia of a legitimate open source project.

This approach has a long tradition in developer culture. Famous examples include the nocode project — a GitHub repository with thousands of stars whose code folder is entirely empty — and the 2016 left-pad incident, in which an 11-line npm package carried an invisible dependency for much of the ecosystem and briefly broke the internet when it was unpublished. These "useless but carefully crafted" projects are fundamentally the tech community turning a mirror on over-engineering and hype. done grafts that tradition precisely onto today's AI tool anxiety, which explains why it resonated so strongly.

Under "Known Limitations," the author writes: "No limitations have been observed to date. Every single request submitted so far has been completed successfully."

After all, if your definition of success is "receiving a Done.", this tool's success rate is eternally 100%.

The Serious Thinking Behind the Absurdity

Jokes aside, this project touches on several issues worth genuine reflection in the AI tool ecosystem.

The Metric Obsession Trap

Token costs are a real pain point for developers. Tokens are the basic unit of text processing for large language models — roughly half an English word or one Chinese character — and mainstream model APIs charge based on total input plus output tokens. As "long context" technology has matured, model context windows have expanded from the original 4K to 100K or even 200K tokens today. But this also means complex agent tasks can easily accumulate vast amounts of redundant information — so-called "context pollution" — causing the model's attention to scatter, response quality to drop, and costs to spiral. Tools that claim to reduce tokens or minimize context pollution address a genuine need.

But done's satire reminds us: any optimization metric divorced from actual output value is meaningless. A tool that drives token usage to near zero, at the cost of doing nothing, makes a joke of the word "efficiency." Genuinely valuable optimization should reduce waste while preserving the quality of completed work — not sacrifice substance for the sake of better-looking numbers.

Collective Fatigue From Tool Overload

"A new plugin appears every few weeks" — that line resonates with a lot of developers. The explosive growth of the AI tool ecosystem has driven real innovation, but it has also created serious information overload and decision fatigue. Huge numbers of tools are functionally near-identical, their marketing copy is overblown, and the ones that actually solve problems are rare. done gives a voice, in the most extreme possible way, to developers who are exhausted from being bombarded with tools.

A Community Turning the Lens on Itself

The author mentioned being inspired by "caveman and co." Projects like these represent a healthy culture of self-deprecating humor within the developer community — using absurdity to process anxiety, using jokes to reflect on trends. They solve no technical problems, but they hold up a mirror and ask whether the industry has gone off course in certain directions.

Conclusion

done is not a productivity tool. It's a carefully crafted piece of technical performance art. With a single Done., it delivers a three-layered critique of AI tool marketing, metric worship, and the proliferation of tools.

In an era when everyone talks about "efficiency gains," "token optimization," and "agent automation," occasionally stopping to laugh — and to ask what we're actually optimizing for, and why — might be more valuable than installing yet another plugin.

After all, as this project proves: the most efficient tool is not necessarily the most useful one.

Done.