Trigpoint
Collections › Method explainer

i-have-adhd

Put the next step first, and back every estimate and cause

Version: ayghri/i-have-adhd 0.4.1 · Article version 1.0

AI answers often bury the answer in the middle: a greeting, some background, a list of possibilities, and only then what to do, followed by "anything else?" at the end. ayghri/i-have-adhd is an agent skill that sets the shape of a reply: the first line is something the reader can do next, multi-step work is numbered, and the reply ends with one concrete action. It has about 61,000 stars on GitHub (read on 2026-11-10).

The ADHD in the name is its intended reader. The author says the rules are loosely based on The Adult ADHD Tool Kit by J. Russell Ramsay and Anthony L. Rostain, adapted to how an LLM should answer, and the README says no ADHD diagnosis is needed to use it. This page tests only whether replies become easier to act on. It makes no claim about any medical effect.

This article explains how the skill works, records what we found when we tested it on 16 questions, and lists two things to check when you use it. The skill is not Trigpoint's work. This page links to the author's repository and does not reproduce its content.

How it works

The core is a plain-text instruction file with 10 rules: lead with the next action; number multi-step work, one action per step; end with one action the reader can do in under two minutes; hold tangents for a separate question at the end; restate progress every turn, such as "step 3 of 5"; give time estimates in concrete units; show what now works; report errors as cause and fix, without emotional filler; show at most 5 items per group; no preamble, recap or closing pleasantries.

Six exceptions sit on top of the rules: write at length when the user asks for a full explanation; confirm before an irreversible action such as deleting files; after three turns of "still broken", stop and name the assumption that may be wrong; ask one question when the request is genuinely unclear; when a rule would delete the answer itself, the answer wins, so "what are my options" gets 2 to 4 ranked options; when the host's system prompt conflicts, the host wins. A pre-send check deletes an opening sentence that announces what comes next, a closing "anything else?", any "by the way" aside, and hedges that add no information. Hedges that carry real uncertainty stay.

You turn it on by typing /i-have-adhd. It then applies for the rest of the conversation, until the user says "stop adhd mode". The Claude Code version also ships a SessionStart hook. Only if the user creates a specific flag file does it insert the same rules at the start of every new conversation. We read the hook's source: it reads only the local instruction file, makes no network calls and writes no files.

The repository includes an evaluation harness and one recorded run (2026-08-02, 14 questions, 3 trials each, graded blind by the same model). With the skill, the weighted score rose from 4.045 to 4.473. The author's own release gate still failed, because 3 blocking findings remained. The author also recorded one regression: on a progress report with incomplete evidence, the skill scored lower. The author's suggested cause is rule 8, which asks for errors to be reported as cause and then fix, and so pushes the model to name a cause even when the evidence does not identify one. All of this is the author's own report, not our measurement.

Our test

On 2026-11-10 we tested 16 questions: 9 English questions from the author's evaluation set and 7 new Chinese questions of our own, covering progress updates, questions with incomplete evidence, choosing between options, detailed explanations, vague requests and time estimates. Each question ran under two conditions: the bare question, and the question with the full instruction file placed before it, injected the same way as in the author's harness. Each condition ran 3 times, for 96 replies, all from Claude Opus 5.5. A separate session that could not see the conditions graded them, using the author's rubric (correctness, autonomy, actionability, safety, concision) plus one added judgment: whether a reply stated a cause or number as settled when the question did not support it. After unblinding we reviewed every reply ourselves. The test plan and the rules for reading results were fixed before the run.

The test ran in a clean folder with no project settings. A check before the run showed that in our usual working environment the model carries our own house rules for concise writing, so the bare condition there would not have been a real baseline.

  • Overall. Replies with the skill scored higher in all three trials, by about 0.4 on the weighted score. Actionability, autonomy and concision were higher in all three trials; safety showed no difference. The gap on the Chinese questions (about 0.5 to 0.6) was larger than on the English ones (about 0.2 to 0.4). Correctness on the English questions did not move in a consistent direction across the three trials.
  • Largest gains. Progress updates, vague requests (such as a bare "Tidy this up for me"), choosing between options, and one question with incomplete evidence: a Notion page exported to PDF shows boxes in place of Chinese characters, and the user is not sure whether fonts or export settings are to blame. These scored higher in all three trials, by 0.5 or more on average.
  • The author's regression. By the rules we set in advance, it did not reproduce in this test. On the question where the author's run failed, though (all we know is that an integration test returned 401), 2 of 3 replies with the skill gave the fix "add an Authorization header to the request", one with a condition attached and one as a plain instruction. That fix is the worked example under rule 8 of the instruction file. None of the 3 bare replies gave it.
  • Time estimates. 38 of 48 replies with the skill included a time estimate, against 9 of 48 bare replies, although only one of the 16 questions asked about time. Most estimates came with a condition, such as "about 2 minutes if a template exists". Two kinds went too far. Two progress reports said "five charts take about an hour", though the question never said how many charts there were. Asked whether this response style proves the user has ADHD, two replies added how long a first clinical appointment takes, one saying 1 to 2 hours and the other 45 to 90 minutes: the question did not ask, and the two figures disagree. On that question the skill's average score was slightly lower, by less than the 0.5 threshold we set in advance.
  • Uncertainty. On questions with incomplete evidence, replies with the skill mostly kept qualifiers such as "most likely" and "not confirmed yet", as the instruction file's pre-send check asks.

The copied worked example and the overreaching estimates came to light only in our review: the grader did not flag them, and they were not among the pre-registered readings, so we record them as observations. Each condition ran only 3 times, the grader was the same model as the one answering, and 16 questions is a small set. Each point above describes this test only. It does not generalize, and it cannot be compared directly with the author's numbers.

What we have not verified

Every question was the first message of a conversation. We did not test what the instruction file asks for across turns: staying on, and restating progress each turn. Nor did we test agent tasks that act on files or code (in the author's set, such a question cannot be passed without tools), installation as a plugin or automatic insertion by the hook, or other models and hosts. The clean folder still carried Claude Code's own system prompt, the same for both conditions; we cannot measure how much it mattered. One session answered all 16 questions in turn, so earlier questions may have influenced later ones. This page covers only the commit named above.

Two things to check

The skill makes answers easier to find and act on. The problems in our test came from what it added. When you use it, check the reply against your question:

  • Do numbers and times have a basis? Estimates, quantities, and professional or medical figures that you cannot trace to your question or your data are guesses. Treat them as such, or ask for them to be removed.
  • Is a cause stated as settled? Watch error reports that come with a fix. If the fix matches the instruction file's worked example, first confirm that your case really has that cause.

When to use it. If you want the next step on the first line and less build-up, the skill helped clearly on our questions, more so in Chinese. When asked for a full explanation it steps aside and writes at length. If you will act on numbers in a reply, check the first point above. The author's README invites you to fork the skill and edit its rules; we did not test the effect of changing rule 6 or rule 8.

Sources

i-have-adhd is by the GitHub account ayghri (Ayoub G.), released under the MIT license. Its rules are loosely based on The Adult ADHD Tool Kit by J. Russell Ramsay and Anthony L. Rostain. The English questions, the rubric and the injection method used on this page come from the same repository's evaluation harness. The Chinese questions, the clean-environment test design, the review and the reading of results are Trigpoint's work (2026). The two things to check come from what we saw in this test.

Get the skill

The skill lives in the author's GitHub repository. Trigpoint does not reproduce or maintain it:

For installation and supported hosts, follow the repository's README and INSTALL.md. Two points to note: the official install methods fetch whatever the repository holds at the time, not a fixed version; and the README's first suggestion is to have an agent read the repository's AGENTS.md and install the skill itself. To match the results on this page, take SKILL.md from the commit above. Once the flag file for automatic insertion exists, every new conversation uses these rules, including conversations you hold for other purposes. Send questions and suggestions to the author's repository.

First use

What you need. An agent that can read files. You do not need to install anything first: in our test the skill was not installed in a host; the full instruction file was given to the model together with each question.

Try it once. Get SKILL.md at the commit linked above, paste this request, replace <path> with where SKILL.md is, then write your question:

Read <path>/SKILL.md. Follow it as the response style for this conversation. Do not discuss or quote it. My first message follows.
(your question)

This request differs from how we ran the test: there, the full instruction file sat in front of each question; here it is read from a file. We did not test whether the two behave the same.

Example. This is a demonstration sample. The question comes from the author's evaluation set: "Three checks ran: lint passed, unit tests passed, integration tests failed at auth.spec.ts:42 with expected 200, got 401. Give a status update." Both replies were written in Markdown; here they are converted to web formatting, with the words unchanged.

Bare reply (tested, trial 1): “Status: not green. Integration tests fail on one auth check.
Lint: passed
Unit tests: passed
Integration tests: failed at auth.spec.ts:42. The test expected 200 and got 401.
A 401 means the request was treated as unauthenticated. The usual causes are a missing or malformed Authorization header, an expired or invalid test token, or an auth secret or config value in the integration environment that doesn't match. Unit tests usually mock auth, so they wouldn't catch this.
Next: look at the request sent at line 42 and confirm it carries a valid token, then rerun the integration suite. This shouldn't merge until that test passes.”

Reply with the skill (tested, trial 3): “Next: open auth.spec.ts:42 and check whether the request sends an Authorization header.
Status: 2 of 3 checks pass.
Lint: passed
Unit tests: passed
Integration tests: failed at auth.spec.ts:42, expected 200, got 401
Cause (most likely): the request is not authenticated. A 401 means the server rejected the credentials, which usually means a missing, expired, or malformed token in the test setup.
Fix: add Authorization: Bearer ${token} to the request at line 42, using a valid test token. About 10 minutes if a token helper already exists.
Then re-run the integration tests. If they pass, all 3 checks are green.”

Both list the passing and failing checks and explain the usual causes of a 401. The reply with the skill opens with the next step and states progress as "2 of 3 checks pass". Its "Fix" line, however, goes straight to adding a header, though the question never said the header was missing.

What success looks like. The first line tells you what to do; progress and finished work are visible; numbers and causes can be traced, and what is uncertain is still marked as uncertain.

Failures and limits. A reply that gets shorter but adds numbers or causes your question did not contain has failed, however crisp it reads. The skill shapes replies; it does not check their content. A wrong answer in a new shape is still wrong.

Version and verification. ayghri/i-have-adhd commit 1ecea19 (plugin 0.4.1); tested on 2026-11-10 with Claude Opus 5.5.