i-have-adhd: The 31K-Star Skill That Stops Claude Code Burying the Answer

A skill with 31.2k GitHub stars on a repo called i-have-adhd is not a joke account — it's currently one of the most-starred prompting skills for coding agents, and it's solving a complaint nearly every Claude Code and Codex user has made at some point: the agent buries the answer under three paragraphs of throat-clearing.

ayghri/i-have-adhd is a small, MIT-licensed instruction file that changes how an agent formats its replies — not what it does. Its tagline states the scope plainly: "ADHD-friendly outputs. No ADHD diagnosis needed!" This lands right after explainx.ai covered the viral "humanising LLM outputs is dumb" debate, which argued style skills like this one shape agent reasoning, not just its prose — a claim worth keeping in mind as you read what follows.

TL;DR — the questions people are actually asking

The before/after that explains why it went viral

The repo's README leads with one comparison, and it's the whole pitch in two paragraphs. Before installing the skill, a typical Claude Code answer about a broken auth flow reads like this:

"Great question! Let me think about this. Your auth flow has a few moving pieces: the middleware, the token verification, and the cookie handling. Looking at src/auth.ts, the verifyToken function (around lines 42-58) seems to be using an older jsonwebtoken API. One approach would be to update the package and rewrite that function. After making the change, you'd want to run the auth tests to confirm nothing breaks. By the way, you might also want to look at your dependency versions overall. Hope this helps! Let me know if you want to dig deeper."

After installing it, the same fix reads like this:

"Run npm install jsonwebtoken@latest, then edit src/auth.ts:42. 1. Open src/auth.ts 2. Replace verifyToken (lines 42-58) with the snippet below 3. Run npm test -- auth.spec.ts Next: paste the first failing line if any test fails."

No greeting, no "one approach would be," no unsolicited tangent about dependency versions, no "hope this helps." Just the command, the numbered steps, and one concrete next action. That's the entire value proposition, and it's why the repo crossed 31k stars — it names a complaint everyone has and fixes it with a single install.

The 10 rules, in order

The skill enforces ten formatting constraints, applied to every response in the conversation once installed:

1. Lead with the next action — the first line is what to do, not context-setting.
2. Number multi-step tasks — no "first... then... after that" prose chains.
3. End with one concrete next step — not an open-ended "let me know if you want more."
4. Suppress tangents — related-but-unasked observations get dropped, not appended.
5. Restate state every turn — the agent doesn't assume you remember where things left off.
6. Specific time estimates — "about 3 minutes," not "a bit" or "shouldn't take long."
7. Make wins visible — passing tests or a clean build get an explicit callout, not silence.
8. Matter-of-fact errors — a failure is reported as a fact, not softened or apologized for.
9. Cap lists at 5 items — long enumerations get trimmed or grouped, not dumped in full.
10. No preamble. No recap. No closers. — no "Great question!", no restating the task back, no "Hope this helps!"

Every one of these is a formatting constraint. None of them touch how the model reasons about the code, which files it edits, or what tools it calls — which is exactly the distinction worth being precise about, and exactly the distinction the humanising LLM outputs debate argued gets muddied in practice, since style instructions sit in the same prompt as the task and can shape the work itself, not just its write-up.

Install it today

Per the repo's own install note, the canonical path is github.com/ayghri/i-have-adhd, with exact per-harness steps documented in the repo's AGENTS.md. For Claude Code, the plugin-marketplace pattern used by comparable community skills looks like this:

claude plugin marketplace add ayghri/i-have-adhd
claude plugin install i-have-adhd

Check the repo's AGENTS.md before running this — plugin marketplace naming can change between releases, and the exact commands for Codex, Cursor, Gemini CLI, OpenCode, Kimi, Qwen, and Antigravity each differ by harness. The README documents each explicitly rather than assuming one install path covers all eight.

Not a clinical tool — and the maintainer says so

This is the part worth being precise about, because the name invites a misread. The repo's tagline is explicit: "ADHD-friendly outputs. No ADHD diagnosis needed!" The credits section goes further:

"Loosely based on The Adult ADHD Tool Kit by J. Russell Ramsay and Anthony L. Rostain. Adapted for how an LLM should respond, not how a human should organize their day."

That's the maintainer drawing the line themselves. The skill borrows organizing principles from a clinical therapy toolkit — action-first thinking, concrete next steps, capped cognitive load — and repurposes them as constraints on model output formatting. It is not diagnostic, not a substitute for an actual accessibility tool, and not marketed as either. It's accurately described as a response-formatting and verbosity-control skill, useful to any developer who wants less narration from their agent, ADHD or not. The localized READMEs (Chinese, Portuguese, Japanese, Vietnamese, Korean, Thai) suggest that broader "I just want less filler" audience is exactly who's actually installing it.

What people are asking: is this a real engineering project or a gist with a catchy name?

The repo structure answers that. This isn't a single SKILL.md gist — it has:

• An evals/ directory with reproducible skill-quality evaluations, including a judge-and-fidelity component from a recent merge — a real test harness for whether the agent actually follows the rules, not just a claim that it does.
• A hooks/ directory tuning SessionStart timeout behavior.
• A tests/ directory, including a recent Windows-specific fix that imports an OpenCode plugin via pathToFileURL — the kind of platform-compatibility bug that only shows up once real users on real operating systems file issues.
• 36 contributors and a merge logged "11 hours ago" at time of writing — this is an actively maintained community project, not a single author's abandoned repo.

That combination — eval harness, hooks, cross-platform test coverage, dozens of contributors, continuous merges — is a stronger maturity signal than the star count alone. Star counts measure discovery; an evals/ directory measures whether anyone is checking the thing still works.

Honest limitations

• It's a prompting-layer skill, not a model fine-tune. Instruction-following reliability is bounded by the underlying model's ability to hold ten formatting constraints across a long session — not something the skill can force. Expect it to hold well on short, focused turns and drift on long or complex ones.
• The eval harness exists because this is a real, open problem. The project shipping its own evals/ directory with a judge-and-fidelity component is a tell: the maintainers know skill-following isn't guaranteed and built infrastructure to measure it, rather than asserting compliance.
• It changes formatting, not correctness. A wrong fix delivered in action-first, numbered format is still a wrong fix — this skill won't catch bad reasoning, it'll just present it faster.
• Cross-harness parity isn't automatic. Claude Code, Codex, Cursor, Gemini CLI, and OpenCode each have different system-prompt architectures, and a skill written primarily against one harness's conventions can behave slightly differently on another — check the per-harness notes in AGENTS.md rather than assuming identical behavior everywhere.
• It can collide with your existing verbosity setup. If you've already turned on Claude Code's built-in Concise output style or installed a compression skill like Caveman, stacking a third formatting layer on top adds instructions without necessarily adding value — test one change at a time so you know what's actually doing the work.

How this compares to what's already out there

i-have-adhd isn't the first attempt at this problem, and it's worth placing next to the two closest things explainx.ai has already covered:

None of these are mutually exclusive, but they're not identical either. Anthropic's Concise setting mostly removes the running "I'll now check X, then Y" narration wrapped around a reply. Caveman goes further, compressing the reply's own surface language into a telegraphic register. i-have-adhd sits in between: it doesn't ask for terser sentences, it asks for a specific structure — numbered steps, a leading action, a capped list length, one closing next-step — which is a different axis than word count. You could reasonably run Concise as your baseline and layer i-have-adhd's structural rules on top for tasks where "what do I do next" matters more than "how many words did that take."

If you want the deeper background on how skills like this get built, versioned, and distributed at all, our complete guide to agent skills covers the SKILL.md format and the trade-offs of always-on instructions versus invoked-on-demand ones. For a broader map of where to find and evaluate skills like this one before installing, see the top 10 agent skill directories.

Bottom line

i-have-adhd earns its star count by naming a specific, common annoyance — agents that bury the fix under a paragraph of throat-clearing — and fixing exactly that, with nothing more claimed than it delivers. It is not a clinical accessibility tool, and the maintainer says so directly. It is a free, MIT-licensed, actively maintained formatting skill that works across most of the coding-agent ecosystem, backed by a real eval harness rather than a bare assertion that it works. Install it, read AGENTS.md for your specific harness, and expect it to hold up better on short turns than long ones — that gap is inherent to prompting-layer skills, not a defect unique to this one.

Related on explainx.ai

• "Humanising LLM Outputs Is Dumb" — the case for rendering at the boundary — the viral HN debate over exactly this class of style skill
• Claude Code ships a Concise output style — Anthropic's own first-party answer to the same complaint
• Caveman skill: token economics and cutting verbose LLM output — a compression-focused alternative
• What are agent skills? A complete guide — how SKILL.md packaging and distribution work
• Top 10 AI agent skills directories & registries — where to find and vet skills like this one
• skills-lock.json: reproducible agent skills — pinning and supply-chain hygiene for installed skills
• Top 10 closed-source and open-source agent harnesses (2026) — the harnesses i-have-adhd installs into
• How to build your first agent skill, step by step — for forking or writing your own

Primary source: github.com/ayghri/i-have-adhd — README, AGENTS.md, evals/, hooks/, and tests/ directories, retrieved September 2026.

Star count, contributor count, and repo structure are accurate as of publication on September 9, 2026 — an actively maintained open-source project can change quickly, so verify current numbers against the repo before citing them elsewhere. Follow @explainx_ai for updates.