When an AI programming assistant gives a long background description but hides the actual command to be executed at the end, the problem usually lies not in the model’s capabilities but in the output structure. i-have-adhd is an output style skill for coding agents: asking the assistant to give the next action first, then adding necessary explanations, and using numbered steps to keep the task status clear.
It is not an ADHD diagnostic or treatment tool and does not provide medical advice. The item name describes a preferred way of presenting information; it can be used even without ADHD to reduce the reading burden of lengthy responses.
Quick Answer
Install in Claude Code:
|
|
After installation, enter /i-have-adhd to enable it in the current session.
Install in Codex:
|
|
After installation, enter $i-have-adhd to enable. This skill will not automatically intervene in all responses by default; it will only change the output style of the current session if explicitly called.
i-have-adhd What changed
This skill is not responsible for writing code, running commands, or modifying projects. It mainly restricts how the Agent organizes answers. Official rules can be summarized into the following categories:
| Rules | Actual results |
|---|---|
| Action first | The first line directly states what to do now |
| Step numbers | Multi-step tasks are listed in the order of execution, without burying actions in paragraphs |
| Recap status of each round | Even if the conversation becomes longer, you can still see where you have been and what the next step is |
| Control branch line | Non-essential background and extension suggestions will not be expanded yet |
| List length limit | Highlight up to five items at a time to reduce the amount of information processed at the same time |
| Clear ending | Leave only a specific next step at the end, not polite words |
It also requires being more specific with time estimates, clearly marking small improvements that have been made, and reporting errors in a plain tone. The core is not to simply “talk less”, but to reorder the information: actions, status, and blocking items first, explanations later.
Install and manage in Claude Code
First add the plug-in market provided by the official warehouse, and then install the plug-in:
|
|
Check the installation results:
|
|
Execute when update is required:
|
|
If you just want to temporarily deactivate rather than delete the plugin:
|
|
To completely uninstall, execute:
|
|
After the installation is complete, enter /i-have-adhd in the session that requires this output method. If the command does not appear in the autocompletion list, restart Claude Code first, as the plugin index is usually read at startup.
Install and manage in Codex
The installation command for Codex is slightly different from Claude Code:
|
|
Check if it is installed:
|
|
When you need to get a new version, you can update the market index and reinstall the plug-in:
|
|
The uninstall command is as follows:
|
|
Entering $i-have-adhd in the conversation will enable this skill. When you want to restore normal output, you can directly tell the Agent stop adhd mode or normal mode. If you still use the old style after switching, it is usually more reliable to create a new session than to repeatedly correct the current context.
How to set the default output style
If you need to enable it manually for each session, you can write persistence rules to the global directive file.
Claude Code uses ~/.claude/CLAUDE.md, Codex uses ~/.codex/AGENTS.md. Add the following:
|
|
Global enablement affects all projects. A safer approach is to manually call it in several sessions to confirm that this rhythm suits your work style, and then decide whether to write the global configuration. For tasks that require didactic explanations, design discussions, or lengthy reasoning, you can also temporarily ask the assistant to fully explain.
Install to other Agent tools
Tools that support the Agent Skills ecosystem can be installed through npx skills:
|
|
Global installation:
|
|
It is also possible to specify the target tool and skip interactive confirmation, for example:
|
|
Check local or global skills using the following command:
|
|
Antigravity users can install directly from the GitHub repository:
|
|
Execute when uninstalling:
|
|
How will an answer change?
Assuming that the build fails, a common answer might first explain the principle of dependency resolution, list several potential reasons, and finally give a check command. With i-have-adhd enabled, the answer will be closer to the following structure:
- First run a read-only command to obtain complete error information.
- Clarify the current status, such as “The build has not passed, and missing dependencies are being confirmed.”
- Select only one most likely repair direction based on the output.
- Run the verification command immediately after repair.
Only one next step is retained at the end, such as “Now send the output of the first command.” This is especially useful for installing dependencies, handling Git conflicts, troubleshooting deployment issues, and tasks that require confirmation back and forth between multiple short steps.
Simplicity does not mean skipping security checks
Official skills reserve exceptions for several types of scenarios:
- When users explicitly request explanations or complete drills, they should still be fully explained.
- Destructive operations such as deleting data and overwriting files must be confirmed first, and security steps cannot be bypassed for the sake of simplicity.
- When there is no progress in multiple rounds of debugging, you should pause and question the original assumptions instead of continuing to stack commands.
- When the conditions are really unclear, you should ask a brief question first to avoid making guesses.
Therefore, this skill is more like a message prioritization rule rather than forcing everything to be compressed. When tasks involve medical, legal, financial or system security, the necessary risk descriptions and validations are still retained.
Customize your own version
If the default rules are too aggressive, you can fork the repository and edit skills/i-have-adhd/SKILL.md. More practical adjustments include:
- Change the list limit from five items to a number that suits you;
- Allow more complete explanations of reasons to be retained in code reviews;
- Specify the use of action-priority structures only for execution-type tasks;
- Add a fixed “current status, risk, next step” template for the team.
Before modifying, it is recommended to read the complete skills document to confirm whether the rules and exceptions are consistent with your work environment. When installing third-party plug-ins, you should also check the repository content, license and subsequent updates, rather than enabling them directly based on the project name.
What is the difference from caveman?
i-have-adhd mainly solves the problem of “what to say first and how to maintain the task status in each round”; caveman puts more emphasis on compressing Agent output, reducing nonsense and token consumption. The two have similar goals, but different focuses: the former favors action sequencing and attentional load, while the latter favors expression compression.
If you just think the answer is too long, try caveman first; if you often can’t find the next step and lose the task status when the conversation becomes long, the structured rules of i-have-adhd are more targeted. Do not overlay too many style constraints at once, otherwise the model may omit necessary information to satisfy the format.
FAQ
Why is there no change after installation?
Plugins are not automatically enabled by default. Claude Code needs to enter /i-have-adhd, Codex needs to enter $i-have-adhd. If you want each session to take effect, write the output rules into the corresponding global command file.
Why can’t I find the slash command or skill?
Restart the Agent first, and then run the plug-in or skill list command to confirm the installation location. When adding a market using a local path, the path should point to the repository root; when using GitHub, use the form owner/repo.
Can it help diagnose or treat ADHD?
No. It only adjusts the output of large language models. It does not assess symptoms, administer medications, and is not a substitute for physician or other professional support. If you are concerned about your medical condition, you should consult a qualified medical professional.
Can the project be used commercially or modified?
The warehouse uses MIT License. Before actual use or redistribution, it is still recommended to check the latest license text and file contents in the repository.
Summary
The value of i-have-adhd does not lie in letting the Agent only answer a few sentences forever, but in stabilizing the order of answers: act first, then explain; explain the status in each round; leaving only a clear next step at the end. For a development process that is often interrupted by long prefaces, too many branches, and vague endings, this is a set of output specifications that can be installed directly or adjusted by yourself after forking.
Project address: ayghri/i-have-adhd