A tool that helps coding agents explore diverse perspectives instead of sticking to their first thought
This tool addresses the issue of AI agents getting stuck on their initial idea and providing only obvious answers. It runs multiple independent thinking frames simultaneously to broaden the range of ideas, then selects the useful ones for deeper exploration. It is particularly effective for tasks where there is no single correct answer, such as design decisions, ambiguous debugging, or API design.
Approaching parallel thinking as a structural problem
Existing chain-of-thought methods tend to anchor on the first sentence, which limits subsequent reasoning. ADHD treats this not as a prompt issue but as a structural problem. It generates N independent agents with different cognitive frameworks to generate ideas without sharing context. A separate critic step then scores the ideas, filters out traps, and deepens the surviving ones. This process mechanically separates the roles of generator and critic, distinguishing them by actual call units rather than promises within a single prompt.
Evaluation results confirmed by independent judges
In comparative evaluations using the same model and problems, ADHD showed a clear advantage over the single-shot approach. Independent LLM judges evaluated six open-ended engineering problems, finding that idea breadth was 1.9 times higher, novelty was 2.9 times higher, and trap detection ability was 5.2 times higher. The ability to filter out 'tempting but broken ideas' that base models often miss is cited as the biggest difference. For example, when handling a problem where a CLI hangs for 90 seconds, it proposed the non-obvious solution that 'the model itself may be unsuitable' rather than a simple retry strategy.
Installation and execution
It automatically detects and installs on approximately 50 or more agent platforms, including Claude Code, Cursor, and Codex. Installation is completed with the single command npx skills add UditAkhourii/adhd. It can be invoked explicitly with the /adhd "problem description" command or runs automatically when it detects an intent to generate ideas. It can also be used as a CLI or TypeScript library, and the exploration range can be adjusted with parameters like framesPerRun or topK. Detailed installation paths and API references are available in the official documentation site and the repository's documentation.
Cost and scope to check before adoption
This tool is released under the MIT license and is already adopted in 17 or more projects, including repowire. However, due to parallel processing and deepening steps, it requires approximately 2.3 times the time and 1.9 times the output volume compared to the single-shot approach. It is suitable for ambiguous problem-solving that requires creativity or interdisciplinary approaches, rather than simple tasks with clear answers. Independent research reviews and benchmark results are transparently tracked in the repository's issues and documentation, so it is recommended to check the costs and benefits for your workflow before adoption.