Vollständiges Transkript anzeigen (1.219 Wörter)
Hi, welcome to another video.
So, there's this new Claude Code plugin that has been blowing up on GitHub recently. It's called i-have-adhd, and it has already crossed about 16,000 stars, which is pretty insane for something that's essentially just a single skill file.
Now, before you ask, no, you don't need an actual ADHD diagnosis to use this. The whole point of this plugin is that it fixes something that annoys literally everyone who uses AI coding assistants, and that's the fact that these models just talk way too much. You ask a simple question, and the AI gives you three paragraphs of preamble, a lecture on best practices, five tangential suggestions you didn't ask for, and then ends with, "Hope this helps. Let me know if you have any questions." And somewhere in the middle of all that buried like treasure is the actual answer you wanted.
So this plugin basically reformats all of Claude's responses to be action-first, numbered, and completely free of filler. And to be honest, after using it for a bit, it's kind of hard to go back. So let's get right into it and see what it actually does, how to install it, and whether it's worth adding to your setup.
Now, let's talk about what this thing actually is. i-have-adhd is a skill for coding agents made by a developer called ayghri. It works with Claude Code as a plugin, and there are also instructions for using it with other agents like Codex. The whole thing is open-source under the MIT license, and it's completely free, which is always great to see.
The interesting part is the thinking behind it. The skill is loosely inspired by an actual book called The ADHD Toolkit, but adapted for how LLMs communicate. The idea is that people with ADHD have limited working memory. They struggle with the gap between understanding something and actually doing it, and they need visible progress to stay motivated. And if you think about it, that's basically everyone when they're tired, distracted, or just trying to ship something at 2:00 a.m.
So the skill takes those principles and turns them into 10 rules that the AI has to follow in every response. Now, let me quickly run you through the main rules, because this is where it gets good.
First, the AI has to lead with the action. So instead of explaining why authentication is complex, it just tells you to run the npm install command and edit the specific file at the specific line. No preamble at all.
Second, every multi-step task gets numbered, and each step is one single bounded action. No, "and then" chains hidden inside a step.
Third, every response has to end with one concrete next action that you can do in under two minutes. This is actually really clever because the hardest part of any task is just starting it.
Then there's tangent suppression, which means the AI finishes the main issue first and only offers secondary stuff as separate follow-ups.
There's also a rule where it restates the state on each turn. So it'll say something like, "Step 3 of 5 done. Schema updated."
Next is backfilling the column so you never lose track of where you are in a long session, which is honestly useful for everyone.
It also has to give specific time estimates, like "about 15 minutes" if tests cover this, and "an afternoon" if not, instead of vague stuff.
Errors get reported matter-of-factly with the cause and the fix. No, "uh oh" or nervous hedging.
Lists are capped at five items maximum, and finally, no preamble, no recaps, and no closing pleasantries. So no more, "let me know if this helps" at the end of every single message.
There are also some sensible overrides built-in. If you ask for an explanation, it'll still explain properly with structure. If something is destructive, it'll confirm first. And if your request is genuinely ambiguous, it'll ask one clarifying question instead of guessing. So it's not just blindly terse, which I appreciate.
Now, let's install it and see it in action. The installation is pretty simple. Run cloud-plugin-install i-have-adhd at i-have-adhd. That's it, you're good to go. Once it's installed, you just open Claude Code and type, "/i-have-adhd" to activate it. And if you ever want to turn it off, you can just run "cloud-plugin-disable i-have-adhd."
So I have Claude Code open here with the skill activated. Let me show you the difference. First, let me ask it something without the skill, just as a baseline. I'll ask how to add JWT authentication to my express app, and you can see it starts with a whole explanation of what JWT is, why authentication matters, the different approaches you could take. It's a wall of text. The actual commands are in there somewhere, but you have to dig for them.
Now, let's try the same thing with the skill active. Let's send it and see. Okay, and look at this. The very first line is the install command, then it gives me four numbered steps, each one a single action with the exact file and line number to edit. At the end, it says the next action is to run the dev server and test the login route, and it estimates the whole thing at about 20 minutes. That's the entire response, no fluff, no lecture, no great question. This is pretty amazing to be honest.
Now, one more thing that's worth mentioning is customization. Since the whole skill is just a single skill.md file, you can fork the repo, edit the file at skills/i-have-adhd/skill.md, and change whatever you want. Maybe you want lists capped at three items instead of five, or you want it even more aggressive with the brevity. You just edit the file, rerun the slash command, and your personalized version is active. The repo has also been translated into a bunch of languages like Chinese, Japanese, Korean, and Portuguese. So that's cool too.
So what's my verdict on this? I think this is one of those ideas that's obvious in hindsight. The models have gotten really good at coding, but the way they communicate is still tuned for politeness rather than usefulness. And this skill fixes that with basically zero cost. It's free, it's open-source, it's a two-minute install, and you can turn it off anytime. If you ask me, the progress restating and the end-with-one-small-next-action rules are the real killers here. Those alone make long agentic sessions way easier to follow. The token savings from cutting all the filler are a nice bonus as well, especially if you're on a limited plan.
There are some downsides, of course. If you're a beginner who actually wants the explanations, this might feel too terse, and you'd need to explicitly ask it to explain things. But for anyone who already knows what they're doing and just wants answers, this is a really good option for sure. Overall, it's pretty cool.
Anyway, let me know your thoughts in the comments. If you liked this video, consider donating through the Super Thanks option or becoming a member by clicking the Join button. Also, give this video a thumbs up and subscribe to my channel. I'll see you in the next one. Until then, bye.
