Why 28,000 developers told their agent they have ADHD

Every coding agent has the same verbal tic. Ask it why your build broke and you get three paragraphs of context, a recap of what you asked, the answer somewhere in the middle, and a closing offer to help further. i-have-adhd, a skill by GitHub user ayghri, attacks that tic with a single markdown file of formatting rules, and developers have responded at a scale that should embarrass the vendors: 28,000 stars and 1,800 forks since May, plus a wave of copycat repos and third-party skill directories listing it.

The pitch is accessibility. The skill tells the agent its reader has ADHD, then enforces ten rules adapted from Ramsay and Rostain's clinical guide The Adult ADHD Tool Kit: lead with the next action, number multi-step work, restate where you are in the process each turn, give time estimates in real units, cap lists at five items, suppress tangents until the current problem is closed, and cut the preamble and the "Hope this helps!" closer entirely. There's an escape hatch for destructive actions, genuine ambiguity, and cases where the user asks for a full explanation.

For developers with ADHD, this is a real accommodation grounded in real literature, and the working-memory argument holds up: if the actionable step scrolls off screen behind context you didn't ask for, the response failed regardless of how correct it was. But the author says plainly that no diagnosis is required, and the star count makes the same point. Most of those 28,000 people aren't filing an accessibility request. They're saying the default register of chat-tuned models is wrong for terminal work, and a repo finally said it out loud.

Why a markdown file beats a system prompt

Anyone who's put "be concise" in a CLAUDE.md or AGENTS.md file knows the instruction decays. The model complies for a few turns, then drifts back to its trained register, because "be concise" is a style preference with nothing behind it. This skill's framing changes the weight of the instruction. "The reader has ADHD and content that scrolls away is lost" gives the model a reason that survives long contexts, the same way models hold onto "the user is colorblind" better than "use fewer colors."

You can call that a prompt hack, and structurally it is one, the benign cousin of every roleplay jailbreak: give the model a human stake and it complies harder. The difference is that the stake is frequently true, and the behavior it produces is what everyone wanted anyway.

Distribution matters as much as the trick. Since Anthropic shipped Agent Skills last October, a SKILL.md file is an installable, forkable artifact rather than a snippet you paste into a dotfile. This repo pushes that portability hard: it ships a Claude Code plugin, a Codex plugin, Cursor skill files, OpenCode support, and extension manifests for Gemini, Qwen, and Kimi, plus always-on hooks so the rules survive session restarts and context compaction. One set of rules, seven agent runtimes. That's the most interesting engineering in the repo, and it's a preview of how agent customization is consolidating around a few markdown conventions the same way editor config consolidated around a handful of file formats.

The evals, read skeptically

Unusually for a prompt repo, there's an evals directory with a rubric, 14 cases, and published results: with the skill applied, a weighted score across correctness, actionability, safety, and concision rose from 4.05 to 4.47, with the biggest gains in concision (+1.14) and actionability (+0.71). Before you quote those numbers, note the methodology: Claude Opus 4.8 generated both conditions and judged them blind, three trials per case, and the results file itself warns that per-case standard deviations reach 0.95. A model grading its own family on a rubric derived from the skill's goals is closer to a smoke test than a benchmark. Credit for publishing the caveats; don't treat the delta as measured truth.

The trade-offs are the part the README undersells. Action-first formatting is the right default when you're driving, but agents earn trust partly through visible reasoning, and a rule set that suppresses tangents will sometimes suppress the tangent that mattered, like the "by the way, this fix assumes your staging config matches prod" aside. The five-item list cap is arbitrary in a way that occasionally fights reality. And matter-of-fact error language is great until you're doing a security review, where you want the agent expansive, paranoid, and slow. The skill's own exceptions clause acknowledges this; whether the model applies the exceptions judiciously is exactly the kind of thing the self-judged evals can't tell you.

What to do with it

Installation is one pasted line into your agent's CLI ("Install the i-have-adhd skill/plugin from the repo URL"), or the Claude Code plugin marketplace, with INSTALL.md covering the rest. The smarter move is the one the author recommends: fork it, edit SKILL.md to your own taste, and install your version. The ten rules are a starting position, not a spec. If you review agent output before shipping it, you may want the reasoning back; if you work in long sessions, the restate-current-state rule alone justifies the install.

The larger signal is for the people building these tools. When a formatting patch written as a disability accommodation becomes one of the most-starred repos of the summer, the vendors' verbosity defaults have failed a much wider audience than the skill's title names. Chat models were tuned for readers who wanted to be walked through things. Agent users want the answer, the next command, and silence. Expect the labs to absorb this into first-party output settings within a couple of release cycles. Until then, the fix is a markdown file, and it's a good one.