Hook
In the quiet corners of GitHub, where stars are the currency of attention, a peculiar repository called "i-have-adhd" crossed the thousand-star mark last month. It is not a model. It is not a framework. It is a set of exactly ten rules—a system prompt—designed to make Claude say less. The first rule: "First line directly provides action." The eighth: "Remove polite openings and repetitive summaries." Reading them, I felt the ghost of an architect staring back from the code. Not the ghost of the AI, but the ghost of every user who ever sighed at a chatbot's three-paragraph greeting before the point. This tiny prompt is not a technical breakthrough. It is a confession. And in the echo of that confession, I hear the unspoken narrative of a billion interactions: we wanted efficiency, but the model gave us etiquette.
Context
Claude Code Skills are custom instructions—essentially system prompts—that modify the behavior of Anthropic's Claude assistant within its code execution environment. They are not fine-tuned models; they are runtime guidance, like a script whispered to a performer before the show. The "i-have-adhd" plugin, created by an anonymous developer, imposes a strict output hygiene: maximum five items, no fluff, action-first. It gained traction not because it solved a novel problem, but because it named a familiar pain. The name itself is a metaphor—borrowing the language of attention deficit disorder to capture a universal cognitive fatigue in the age of information overload. In my years analyzing DeFi protocols during the 2020 summer, I saw a similar pattern: complex systems (like Uniswap's liquidity pools) were often misused not because they were technically flawed, but because the user interface (the docs, the prompts) assumed a level of patience that the market did not have. The same applies here. The AI output is a UI; when it is verbose, it loses users.
Core: The Narrative Mechanism and Sentiment Analysis
The true power of this plugin is not in its rules but in its implicit indictment of the default. Claude's native output, like most large language models, is polite, thorough, and repetitive. It begins with "Sure, I can help with that" and ends with "Let me know if you need further assistance." For a user typing a command like "list open TCP ports," this is noise. The plugin removes that noise. But the narrative here is deeper: it signals a shift in the implicit contract between user and AI. The default contract was "I am a helpful assistant who provides comprehensive answers." The new contract is "I am a task-completing agent who gives you what you asked for, only what you asked for, in the fewest words possible." This is the rise of Task-Oriented AI as a dominant narrative, competing with the older Comprehensive Assistant narrative.
From my experience debugging smart contracts for Project Aether in Zurich—where a five-line reentrancy hole could drain millions—I learned that the most dangerous flaw is often the one that is never questioned. Here, the flaw is the AI's default verbosity. The market has spoken: over a thousand developers voted with their stars that they find the default obstructive. Sentiment analysis of the repository's comments reveals a common emotional thread—exhaustion. Words like "overwhelmed," "too long," "just get to the point" appear repeatedly. This is not a feature request; it is a cry for cognitive economy.
But let's apply my original analysis framework. The emotional tone here is not anger; it is melancholic clarity. Users are not upset at Claude; they are disappointed that the model's intelligence is buried under unnecessary scaffolding. In the same way that a liquidity pool that closes due to a bad governance vote leaves behind only the intent of its creators, here, when the fluff is stripped, only the intent of the question remains. When the pool empties, only the intent remains.
Contrarian: The Blind Spots of Brevity
Yet, I am skeptical of any one-size-fits-all cure. The same developers who applaud this plugin may, a month later, discover its shadow. Forcing brevity on every output strips context. In my work as a Web3 research partner, I often need Claude to walk me through the code of a failed NFT marketplace to understand the exact sequence of its exploit. If the output is limited to five bullet points, I lose the intermediate steps—the very steps that hold the narrative of the hack. Brevity is a privilege of the informed; for the novice, it is a trap. This plugin, like many well-intentioned tools, assumes the user knows what to ask. In reality, the most valuable AI interactions are often exploratory, where the model's verbose scaffolding helps the user discover the unknown unknown.
Moreover, there is an ethical dimension that the community's euphoria has obscured. The name "i-have-adhd" appropriates a clinical term to market a utility. It is evocative, yes, but it also trivializes a real condition. In my NFT community days in London, I saw how easily a narrative can be co-opted for virality, losing its original sincerity in the process. The plugin's success may unintentionally signal that neurodivergence is merely a metaphor for efficiency, a dangerous simplification. The audit is not a check; it is a confession. This plugin confesses that the industry still does not know how to design for diverse cognitive needs—it simply applies a blunt instrument called "shorten it."
Takeaway: The Next Narrative
The "i-have-adhd" plugin is not the endgame; it is the crack in the facade. The next narrative will not be about shorter prompts but about dynamic output orchestration—AI that adapts its verbosity to the task, the user's expertise, and the context. Imagine a system that knows when to give a ten-page report or a one-line command. The ghost in the prompt is not the developer who wrote these ten rules; it is the architect of the original model, who forgot to ask the user's permission before speaking. The question every builder must now answer is not "Can my model reason?" but "Can my model shut up when asked?" In that silence, the real value emerges.
