A GitHub project called i-have-adhd hit 22.8K Stars—not with a smarter model, but with a SKILL.md that teaches AI to shut up. It surfaces an uncomfortable fact: AI is hard to use, often not because it's dumb, but because it talks too much.

What this is

i-have-adhd is an open-source project on GitHub with roughly 22.8K Stars and 1.5K Forks at press time, under the MIT license. It doesn't train models. Instead, it feeds a SKILL.md to Coding Agents (Codex, Claude Code, Gemini CLI, Copilot, Qwen Code, and others) instructing them to lead with answers and skip the filler when editing code.

The core rules come down to 10 items, in four groups:

  • Lead with action: Put the answer or the next command on the very first line; code snippets come before explanations.
  • Report progress: Multi-step tasks must be numbered, and every turn restates how many steps are done.
  • Be specific: Error reports must include file location, cause, and the fix action; time estimates need concrete units.
  • Cut noise: Lists stay under 5 items; no "Great question" filler.

At its core, it translates the "do first, explain later" communication template commonly used by ADHD patients into rules an LLM can read.

Industry view

Supporters see this as prompt engineering going product-grade—the same Claude 4 with and without these rules can deliver developer experiences a full order of magnitude apart.

Critics are more direct: if a model needs a SKILL.md to stop padding its output, wasn't base-model alignment undercooked? This pushes a product-layer problem onto users. Others add that output rules govern only how things are said, not whether they're correct—an Agent that fires off a wrong command first can be more dangerous than one that pauses to explain. Claude Code and Codex also differ substantially in framework, and there's no systematic comparison showing the same SKILL.md actually works on both ends.

Impact on regular people

  • For enterprise IT: Have dev teams trial Skills like this—a low-cost way to validate whether you can cut wasted back-and-forth rounds.
  • For individual careers: Programmers' core value is shifting from "writing prompts" to "defining the structure of a good answer." Knowing what to say first and what to cut now matters more than tuning parameters.
  • For consumer markets: Ordinary users won't touch this directly yet, but it carries a clear reminder: AI product UX is far from settled, and "the smartest AI" and "the most usable AI" are two very different things.