i-have-adhd Skill Review: 7,201 Bytes Telling Claude to Stop Burying the Answer
The whole skill is 7,201 bytes, and its first rule fits in one sentence: "The first line is something the reader can do." ayghri/i-have-adhd has 58,478 GitHub stars and 3,383 forks, and it was created on 2026-05-13. Hacker News search shows four submissions of it: three scored 2 points each, and the fourth, on 2026-09-08, reached 542.
I read the skill directory and the Claude Code hook files from a shallow clone and ran none of it. Nothing was run through any API. Token figures below are file bytes divided by 4, not measurements.
TL;DR / Key Takeaways
- The i-have-adhd
SKILL.mdis 7,201 bytes (420 B frontmatter plus 6,781 B body); at bytes ÷ 4 that is an estimated 1,800 tokens, an estimate from file size and not a measured session figure. - The ayghri/i-have-adhd repository has 58,478 stars, 3,383 forks, an MIT license and release v0.4.1 (2026-10-19); it holds one skill, though GitHub code search counts 2
SKILL.mdfiles because a same-size Cursor copy exists. - The skill's frontmatter sets
disable-model-invocation: true, so it is written to load only when you type/i-have-adhd, and it tells the agent to stay on until you say "stop adhd mode". - The maintainer's own
evals/RESULTS.md(dated 2026-08-02, 14 cases, 3 trials) scores 4.473 with the skill against 4.045 without, a +0.427 weighted gain, and also records "Release gate: FAILED" because 3 blocking findings remained. - The 542-point Hacker News thread had 72 top-level comments; one commenter reported that Claude kept the style for "a few turns at most", which is the exact behavior the skill's persistence rule says should not happen.
What the skill actually says
It's an output-style skill with ten rules (lead with the next action, number steps, end with one next step, restate state every turn, no preambles or closers, and so on), six conditions for breaking them, and a pre-send checklist. Two lines carry the design, plus the persistence promise quoted below:
"The first line is something the reader can do. Not context. Not a plan. The action."
"State cause and fix." (rule 8, on error reports)
The frontmatter sets disable-model-invocation: true and agents/openai.yaml sets allow_implicit_invocation: false, so by the files' own settings it is on-demand: type /i-have-adhd, and it stays on until you say "stop adhd mode". Upkeep is active: 16 commits to SKILL.md, the latest on 2026-09-15, including three rewordings of rule 9 (the five-item cap) in two days.
What it costs you in context
Arithmetic from file size, divisor 4 bytes per token. Real tokenizer counts will differ, and none of this was measured in a live session.
| File | Bytes | Est. tokens (bytes ÷ 4) | Role |
|---|---|---|---|
skills/i-have-adhd/SKILL.md | 7,201 | ~1,800 | The skill; 420 B frontmatter (~105) plus 6,781 B body (~1,695) |
.cursor/skills/i-have-adhd/SKILL.md | 7,201 | ~1,800 | Cursor copy, same size |
skills/i-have-adhd/agents/gemini.toml | 2,157 | ~539 | Gemini CLI custom command with a condensed 10-rule prompt |
skills/i-have-adhd/agents/openai.yaml | 238 | ~60 | Display name and allow_implicit_invocation: false |
hooks/always-on.mjs (+ hooks.json, 549 B) | 1,766 | 0 (runs outside context) | Opt-in SessionStart hook for startup, resume, clear, compact; injects the SKILL.md body |
evals/RESULTS.md | 4,915 | 0 (never loaded) | First recorded eval run |
On demand, SKILL.md alone is the loaded set: about 1,800 estimated tokens. With the always-on hook, the stripped body plus a short banner is injected at startup, resume, clear and compact, about 1,700 estimated tokens each time. Whether this skill makes a whole session cost less or more is unpublished; the maintainer says reducing token usage "is not a target".
What the community reports
Discussion sits in one Hacker News thread and the repo tracker (31 open issues, 49 closed; 73 merged pull requests, 53 open). I found no blog write-ups worth quoting.
| Source | Date | What was reported | Link |
|---|---|---|---|
| Hacker News thread, 542 points, 72 top-level comments | 2026-09-08 | Mostly venting about Claude's verbosity and unrequested "I did not edit X" narration (e.g. ryandrake, jp57), plus doubts that a skill is the right tool. | item 49610631 |
| HN comment by sleazebreeze | 2026-09-08 | With this skill, Claude models "only maintain the conciseness for a few turns at most", even with a CLAUDE.md instruction and manual repeats. "I don't think we can skill our way out of this one." | comment 49611957 |
| HN comment by hbn | 2026-09-08 | Objected to the README's install method (paste a prompt telling the agent to install from the repo), fearing it could land on a squatted, similarly named repo. | comment 49614192 |
| HN comment by ayghri (maintainer) | 2026-09-11 | Says it "started as a joke", suits work you already understand, is "not good for exploration and learning", and that reducing token usage "is not a target". | comment 49652042 |
| GitHub issue #263 (open) | 2026-10-18 | The always-on hook's Node and OpenCode copies leave an empty frontmatter block in the injected ruleset. Contributor SU-AN-coder says a fix branch passes 80 tests but PR creation was refused for their account. | issue 263 |
| GitHub issue #267 (open, 0 comments) | 2026-10-21 | joaobcjr, from real use in the Claude desktop app for macOS, proposes a rule to put anything the reader must copy into a code block. | issue 267 |
The sharpest dissent is whether a skill can do this at all. Replies to sleazebreeze split: soontimes said a stop-event hook works but sometimes over-compresses, and alwillis pointed to Claude Code's built-in Concise output style (untested by me). iambenm answered "why a skill, not AGENTS.md" by saying things in AGENTS.md "tend to lose weight as context grows". Separately, egeozcan, who says they have ADHD, objected to the name, a branding complaint rather than a functional one.
What the maintainer's evals show
evals/RESULTS.md compares the bare prompt with the skill body injected: 14 cases, 3 trials, 84 graded rows, judged blind by the same model family that generated them. Weighted score: 4.045 baseline, 4.473 with the skill. Concision moved most (+1.143), then actionability (+0.714). The skill won 10 cases, tied 2, lost 2.
The file then says "Release gate: FAILED": 3 blocking findings remain against 7 for the baseline, and the rule is absolute. Two come from a case no run can pass because tools were disabled. The real regression is partial-success at −0.63, where the grader noted the response "asserts 'missing auth header' as the definitive cause and prescribes a specific fix without any evidence". The file ties that to rule 8.
How to install it
Verified against INSTALL.md in the clone I read (commit 1ecea19, dated 2026-10-11):
# Claude Code
claude plugin marketplace add ayghri/i-have-adhd
claude plugin install i-have-adhd@i-have-adhd
# Optional: load the ruleset automatically at session start
touch ~/.claude/.i-have-adhd-always
# Undo
rm ~/.claude/.i-have-adhd-alwaysThen type /i-have-adhd. The README's headline method is to paste a prompt asking your agent to install from the repo; given hbn's objection, use the explicit commands.
Where it earns its keep, and where it doesn't
It earns its keep as a toggle. When a session drifts into wall-of-text answers, a roughly 1,800-estimated-token switch you flip on and off is a reasonable thing to try. The rules are concrete, and the maintainer's own scope, work "you already understand", is the honest use case.
Now the criticism.
- Persistence is unproven. The skill says its rules "do not expire"; the one detailed user report says they faded within a few turns. The shipped hook fires on SessionStart only, so it re-injects after compaction but not on each response. The six-turn persistence scenario under
evals/scenariosis a capture recipe whose write-up says it doesn't establish compaction handling, and I found no published scores for it. - The evals are thin and dated. Three trials, a same-family judge, model-graded dimensions, and no data on whether readers with ADHD do better. RESULTS.md is dated 2026-08-02;
SKILL.mdhas had at least six commits since. - Rule 8 can push the model to guess. The maintainer's own eval found "cause and fix" pressured a confident cause without evidence.
- The headline install method is a prompt, the pattern hbn flagged.
Skip it for long exploratory sessions where you want the model to teach you; by the maintainer's account, that is where it is weakest. And if a built-in concise style already fits your workflow, as alwillis suggested, 1,800 estimated tokens may not buy much.
FAQ
What does the i-have-adhd skill do?
It is an output-style skill with ten rules: lead with the next action, number multi-step work, end with one next step, and drop preambles and closers. SKILL.md is 7,201 bytes, MIT licensed. Turn it on with /i-have-adhd, off with "stop adhd mode".
How much context does the i-have-adhd skill use?
SKILL.md is 7,201 bytes, so bytes ÷ 4 gives an estimated 1,800 tokens, or about 1,695 for the 6,781-byte body. These are estimates from file size, not measured session figures.
Does i-have-adhd stay on through a long session?
The skill text says its rules apply for the rest of the session, but one Hacker News commenter reported that Claude models kept the style for only "a few turns at most". The optional always-on hook re-injects the ruleset on startup, resume, clear and compact, not on every response, and the repository publishes no persistence scores.
How do I install i-have-adhd in Claude Code?
Run claude plugin marketplace add ayghri/i-have-adhd, then claude plugin install i-have-adhd@i-have-adhd, as listed in the repo's INSTALL.md. For always-on mode, create the flag file with touch ~/.claude/.i-have-adhd-always.
Skill by Ayoub G. (ayghri), MIT licensed per GitHub, the README and the SKILL.md frontmatter. Source: github.com/ayghri/i-have-adhd. Quoted lines are short excerpts for review purposes. Reviewed October 24, 2026 from source, GitHub data and public Hacker News threads; nothing was run through any API. Community reception shifts over time, so check the issue tracker for current status.