uditakhourii/adhd

ADHD is a skill for AI coding agents that addresses a specific problem with how AI models reason through open-ended questions. When an AI works through a problem step by step in a single chain of thought, it tends to anchor on whatever direction it explores first, and subsequent ideas stay close to that initial framing. The ADHD skill tries to break that pattern by running multiple separate, isolated reasoning processes in parallel, each starting from a deliberately different perspective or cognitive frame, before a second pass evaluates and compares the results. The two-phase process works like this. In the first phase, the skill picks several cognitive frames, such as approaching a problem from a cost perspective, a security perspective, or an unconventional angle. It spawns separate AI agent calls for each frame, and each call is prevented from seeing what the others produced, so the ideas stay independent. In the second phase, a separate critic call scores every idea across novelty, viability, and fit, identifies ideas that seem appealing but are likely to fail, clusters the remaining ideas by underlying approach, and expands the top results into more detailed sketches with risks and first steps. The skill is designed for problems where the goal is to generate options rather than execute a known plan: design decisions, naming things, choosing an architecture, debugging problems that do not have an obvious cause, or any question of the form "give me a few ways to approach this." It is not meant for tasks with a clear correct answer. Installing it takes one command that auto-detects which AI coding tool you are using. It works with Claude Code, Codex, Cursor, Windsurf, and around fifty other tools. After installation you can invoke it explicitly by name with your problem as input, or it can trigger automatically when you ask an open-ended ideation question. The README includes benchmark results from testing across six open-ended engineering problems, showing large improvements in idea breadth, novelty, and trap detection compared to a single-shot baseline. The full README is longer than what was shown.