Hook
Over the past seven days, a GitHub repository called “i-have-adhd” has accumulated 1,100 stars and zero pull requests. Its entire value proposition is a list of ten rules designed to make Claude’s output shorter. No model fine-tuning. No new architecture. No cryptographic breakthrough. Yet inside the crypto development community—where every second of AI-assisted debugging can mean the difference between a shipped contract and a rekt protocol—this plugin has become a silent standard. The ledger of GitHub stars tells a story that the hype around AI capabilities often forgets: utility is defined by output, not model size.
Context
Claude Code is Anthropic’s code-execution environment, widely used by blockchain developers to write smart contracts, audit Solidity, and scrape on-chain data. By default, Claude produces verbose, polite responses—complete with disclaimers, summaries, and confirmations. For a task-driven user fixing a reentrancy bug or querying a Uniswap pool, that verbosity is friction. “i-have-adhd” is a community-created skill that injects a system prompt instructing Claude to: skip greetings, deliver the first line as action, limit lists to five items, avoid redundant conclusions, and omit any information not directly requested. The developer, anonymous, named it after their own attention span. The repository’s README is blunt: “This does not make Claude smarter. It makes it shut up and work.”
From my 2018 audit of EtherCity’s off-chain ownership records to my 2022 exposé on NFT wash trading, I have learned that the most dangerous failures are the ones hidden in plain sight—in this case, the belief that more AI output is better. The crypto industry, built on haste and hype, is particularly vulnerable to this fallacy. The “i-have-adhd” project is a counter-signal worth dissecting.
Core: The Systematic Teardown of AI Output Overhead
The plugin’s technical DNA is pure Prompt Engineering. There is no model-wide weight change, no reinforcement learning, no data poisoning. It is a set of inference-time output constraints applied at the system prompt level. Ten rules, each targeting a specific friction point in the human-AI interaction loop:
- First line is the action. Eliminates the “Sure, I can help you with that” opening. For a developer auditing a reentrancy guard, this saves 200 milliseconds per query—trivial alone, transformative across a day.
- No polite framing. “I’d be happy to explain” becomes a direct code snippet. The cost: social nuance. The gain: task completion speed.
- Max five items. Respects Miller’s Law—human working memory holds 7±2 chunks. Crypto developers juggle RPC endpoints, gas calculations, and contract state; reducing cognitive load is a measurable efficiency gain.
- No repetitive summary. “In conclusion…” is banned unless explicitly requested. Claude’s default behavior often repeats the solution in a new paragraph. This rule kills that redundancy.
- Remove all greetings and sign-offs. No “Hello, I’m Claude.” No “Let me know if you need anything else.” The conversation ends when the task ends.
- Only keep current relevant info. If the user asks for a Solidity function, Claude does not include background on the ERC-721 standard unless asked. This is especially critical in crypto, where context switching between protocols (Uniswap vs. Compound) can introduce bugs.
- No self-awareness disclaimers. “I am an AI, please verify” is suppressed. The developer is already aware. The disclaimer adds zero value and undermines trust in the output.
- Use bullet points for code suggestions. Structural clarity over paragraph flow.
- Limit to one example unless asked. Prevents the AI from generating three variations of the same function when only one is needed.
- No apology loops. If Claude makes an error, it corrects without saying “I apologize, you are correct.” Time saved.
Individually, each rule seems trivial. Collectively, they represent a hostile takeover of Claude’s output grammar. The core insight is that the bottleneck in AI-augmented crypto development is no longer reasoning quality—it is output density. Models like Claude 3.5 can already solve complex audit questions. But they wrap those solutions in layers of politeness, repetition, and caution that slow down the expert user. The “i-have-adhd” plugin surgically removes those layers.
From my experience investigating the Curve Finance governance centralization in 2021, I saw how a small group of whales controlled 60% of protocol decisions through voting weight. Similarly, here a small set of ten rules controls 100% of Claude’s output format. The power is not in the rules themselves, but in the behavior they enforce. The project is a textbook case of what I call “output-level governance”—a subtle but decisive influence on how AI decisions are communicated to humans.
Yet the plugin also exposes a deeper flaw: Anthropic’s default system prompt is not optimized for task-oriented users. The very existence of this project is a confession that Claude’s default output is bloated. Silence in the code is the loudest confession—Anthropic has not addressed this, leaving the community to patch the interaction layer. This mirrors the crypto industry’s pattern of letting users bear the cost of incomplete product design.
Contrarian Angle: What the Bulls Got Right
Critics will argue that stripping disclaimers and politeness makes Claude less safe. They have a point. In a smart contract audit, the AI’s caution to “verify with a human” might be the only thing preventing a developer from blindly deploying a flawed function. The plugin’s rule 7 suppresses that safeguard. For a novice developer, that could be catastrophic.
But the bullish counter-argument is equally valid: the target user is not a novice. The plugin’s 1,100 GitHub stars likely come from experienced crypto developers who already know to double-check critical code. For them, the AI’s hand-holding is noise, not signal. The bull case is that precision of communication is a form of respect for the user’s expertise. The plugin treats the developer as a competent peer, not a student. In a field where speed is capital, that respect translates into measurable productivity gains.
Moreover, the plugin does not remove safety entirely—it removes the illusion of safety. A developer who pastes a function without verifying it is dangerous regardless of disclaimers. The plugin forces the developer to take ownership of the output. I do not cover the story; I follow the code—and the code here is a prompt that says: trust yourself, not the AI’s warnings.
The bulls also correctly note that this plugin is a proof of concept for a broader trend: the commoditization of output style. Just as crypto has flash loans and MEV, AI will have shorthand agents and minimalist prompts. The “i-have-adhd” ruleset is the first draft of a new genre—call it “developer-centric AI output.” It is not the final word, but it sets a direction.
Takeaway: The Real Signal Behind the Stars
The “i-have-adhd” project will likely be absorbed into Anthropic’s official system prompt within three months. The creator should negotiate an integration deal now, or watch their utility vanish before the mint even cools. But the deeper signal for the crypto industry is this: as AI models become commodity intelligence, the competitive advantage will shift to how they are instructed to communicate. The next killer app in crypto development may not be a faster L2 or a new consensus mechanism—it may be a ten-line system prompt that makes an existing model actually useful.
We traded verbose outputs for concise actions. That trade is worth making, but only if we remember who is left out. The plugin’s name—i-have-adhd—is a clever marketing hook, but it also frames attention deficit as a problem to be solved by silencing the AI. What about the developers who need context, who want explanations, who learn from the AI’s digressions? The plugin is a tool, not a universal fix.
The ledger remembers what the hype forgets: the real efficiency gain is not in the code, but in the silence between the lines. Use it wisely.