i-have-adhd: The Skill Making AI Coding Agents Blunt

Three weeks ago, a machine-learning researcher pushed a folder of markdown files to GitHub. Today that repo sits at #1 on GitHub's daily trending leaderboard, ahead of every funded AI startup and dev-tool launch on the page, and picked up 3,882 stars in the last 24 hours alone — on top of the 38,226 it already had.

The project is i-have-adhd, and it doesn't add a single feature to your AI coding agent. It just tells the agent to stop burying the point.

What i-have-adhd actually is

i-have-adhd is a free, MIT-licensed "skill" — a set of instructions an AI coding agent reads before it answers you — built by Ayoub Ghriss, who lists himself as working with the CLIMB research group at the University of Colorado Boulder. Per the repo's own description, it's "A skill to stop your coding agent from burying the answer. ADHD-friendly output." The README's own tagline is more direct: "ADHD-friendly outputs. No ADHD diagnosis needed!"

There's no model behind it, no API, no dashboard. It's a prompt-engineering skill file that any agent supporting Claude's skill format — or an equivalent plugin format — can load, and it rewrites the shape of every response that comes after.

The mechanism: ten rules that reorder the answer

Per the project's SKILL.md, the file opens with five stated "facts" about ADHD and working memory, then ten rules translating those facts into instructions for how a language model should structure text. The gist of the ten rules:

1. Lead with the action — a command, path, or snippet — before any explanation.
2. Number multi-step tasks, one bounded action per step.
3. End every turn with exactly one thing that's doable in under two minutes.
4. Finish the current issue before mentioning a secondary one.
5. Restate progress each turn (e.g., "step 3 of 5 done") instead of assuming memory.
6. Use concrete time estimates instead of vague ones.
7. Show what now works, in concrete terms, instead of a recap.
8. State errors and fixes plainly, without hedging language.
9. Cap lists at five items, grouped by relevance.
10. Skip the preamble, the recap, and the sign-off.

None of that requires a new model or a fine-tune. It's a formatting contract layered on top of whatever agent you're already running, which is exactly why it was cheap enough for one person to ship and fast enough for thousands of people to try in three weeks.

Why it's the thing everyone's installing this week

The growth curve is the actual story here, and it's visible on GitHub's trending page in real time: 38,226 stars and 2.2k forks as of today, with 3,882 of those stars added in the past 24 hours alone. The project's commit history shows activity going back to at least August 21, 2026, meaning most of that growth happened in under three weeks.

It's not a one-person hobby project anymore, either. The repo has 61 merged pull requests and 18 open issues as of this writing — contributors adding install docs for new tools, fixing platform-specific bugs, and arguing over the exact wording of individual rules — one merged PR is titled "rules: reword rule 9 to presentation-only ranking".

It works across almost every AI coding tool, not just one

The reason it spread this fast is partly distribution: it isn't locked to a single agent. Per the project's INSTALL.md, it ships install instructions for 15 different tools — Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot (including the VS Code extension and Copilot CLI), Zed, OpenCode, Qwen Code, Amp, Antigravity, AstronClaw, Hermes, Kimi Code CLI, Pi, and Oh My Pi. That's a wider net than most single-vendor launches cast — the kind of cross-tool bet last month's crop of agent-first browsers also made, just aimed at output formatting instead of browsing.

The activation behavior isn't uniform across those tools, and the project's own install docs are upfront about the inconsistency: in Claude Code, Qwen Code, and Codex, nothing happens until you invoke the skill explicitly (typing /i-have-adhd turns it on for the session), while some other integrations load the ruleset automatically at session start. That's a rough edge for anyone expecting one install step to behave the same way everywhere — check the specific instructions for your tool before assuming it's live.

What's still missing

Three weeks in, there's no versioned release yet — the releases tab is empty, so anyone installing it today is pulling straight from main. There's no independent study behind the "ADHD-friendly" framing — just the five stated "facts" in SKILL.md and the maintainer's own design notes; this is one developer's interpretation of ADHD-related communication advice applied to LLM output, not a clinically validated intervention, and the project doesn't claim otherwise. And because the whole thing is a markdown file you can fork, per the README the intended customization path is literally editing SKILL.md yourself — which means the "rules" every user gets are whatever that person's fork says, not a single maintained standard.

The takeaway

i-have-adhd's growth curve is a clean data point for a pattern this blog has tracked before: a free, single-purpose GitHub repo with a sharp README can now out-compete funded product launches for attention within days, the same distribution shift that put God's Eye View at #1 on trending in four days last month. The mechanism here is even smaller than that one — it's not new data or a new UI, it's ten formatting rules — which makes the lesson sharper: the bottleneck for a lot of AI tools right now isn't capability, it's whether the answer is usable the moment it lands. If you're testing this yourself, start with one tool you already use daily, install it following the exact steps for that tool in INSTALL.md, and see whether the shorter, front-loaded answers actually save you time before you fork the rules to taste.