Hook
Over the past seven days, a single GitHub repository has quietly accumulated 1,100 stars. It's not a new DeFi protocol, an L2 scaling solution, or a zk-proof library. It's a 10-line Claude Code skill called 'i-have-adhd.' And the fact that it went viral tells me something deeply unsettling about the state of AI tools in blockchain development.
We built these models to think for us. To audit our code, to explain complex tokenomics, to draft governance proposals. But somewhere along the way, the output became a wall of text—polite, thorough, and utterly exhausting to read. The ADHD skill isn't a novel breakthrough; it's a cry for help from developers drowning in verbosity.
Context
The 'i-have-adhd' skill is a set of system prompts designed for Anthropic's Claude Code—a terminal-based AI assistant for coding. It does not enhance Claude's reasoning capabilities. It does not fine-tune the model. It is pure prompt engineering: a collection of ten rules that force the AI to output with ruthless efficiency. Rules like 'first line provides the action,' 'no polite openings,' 'list no more than five items,' and 'remove all conclusions and summaries.'
For the blockchain community, this is not a trivial toy. Every week, I see projects integrate Claude or GPT-4 into their smart contract development pipelines—for code generation, audit assistance, or documentation. The AI writes verbose responses that contain the answer within paragraphs of unnecessary context. Developers scroll past the politeness to find the actual code, wasting cognitive cycles.
Based on my own audit experience—I've manually reviewed over forty ICO whitepapers and three failed DeFi projects' tokenomics during the 2017 bull run—the single largest friction point for adopting AI in blockchain development has never been accuracy. It has been format. The models are smart, but they don't respect the developer's attention span.
Core
The ADHD skill's popularity exposes a fundamental misalignment between AI output design and the workflow of blockchain engineers. Let me break down why this matters beyond a simple GitHub trend.
First, the rules map directly to principles of cognitive load reduction. Miller's Law states that human working memory can hold roughly seven items—the ADHD skill limits lists to five. Rule number four: 'Only keep the information needed at the current stage.' This mirrors how we read smart contracts: we don't need the full context of a reentrancy guard when we're just looking at the withdraw function signature. We need the specific check and the revert condition.
Second, the skill reveals that Claude's default behavior—and by extension, many blockchain AI assistants—prioritizes being helpful over being efficient. The default response often includes an introduction ('Sure, here's how you can implement a Merkle proof...'), a step-by-step walkthrough, and a concluding summary. For a developer who already knows what a Merkle proof is, that's three layers of noise. The task-oriented user wants the code first, the explanation later.
Third, the skill's name—'i-have-adhd'—is a masterstroke of emotional branding. It signals that this tool is for people who feel overwhelmed by information overload. In a space where even a simple Uniswap swap can generate 300 lines of Solidity, where governance forums are filled with multi-page proposals, the promise of brevity is seductive. It's not about actual ADHD; it's about respecting the finite attention of a developer who has four terminals open and a mainnet transaction pending.
From my analysis of the rules, I see a hidden philosophy: the model should behave like a command line, not a conversation. This is a fundamental design choice. Conversations require turn-taking, context building, and politeness. Commands require a single, precise action with minimal decoration. The blockchain industry runs on commands—eth_call, cast send, forge test. We don't say 'please' to the node.
Contrarian
Yet, I must play pragmatist. The ADHD skill, for all its cleverness, introduces a dangerous trade-off: the risk of over-optimization.
In my investigation of algorithmic stablecoin failures during the 2020 DeFi summer, I interviewed twelve users who lost savings due to oracle failures. One common thread was that the smart contracts often contained warnings in comments—'this function is safe only if the oracle price is within 5% of the EMA.' A model forced to be brief might strip those warnings as 'unnecessary.' The rule that says 'remove all conclusions' could easily suppress a critical security summary.
Is a perfectly concise AI response worth missing a subtle vulnerability flag? I don't think so. And yet, the market sentiment suggests otherwise. The 1,100 stars indicate a strong demand for brevity. The real insight is that we need an intelligent layer that decides when to be verbose and when to be terse based on the task's risk profile, not a blanket set of rules. The ADHD skill is a blunt instrument; it is a signal of a problem, not its final solution.
Furthermore, this trend could push AI assistants to optimize for developer satisfaction metrics—like 'time to first code snippet'—over code safety. If an AI knows it will be penalized for long outputs, it might skip crucial context. Faith in the protocol is not faith in the people. We must ensure that brevity does not become an excuse for negligence.
Takeaway
The 'i-have-adhd' skill is a canary in the coal mine for blockchain AI tools. It tells us that the next frontier isn't smarter models—it's better interfaces. The industry is moving from 'can the AI find the reentrancy bug?' to 'can the AI tell me the reentrancy bug in one line, without the fluff?' The temple of decentralization was built on code, but we forgot that code must be read by humans. This skill reminds us that truth is not a token you can trade—it's a signal that must be delivered without noise. The question is not whether we can make AI shorter, but whether we can make it wiser.