alt Hacker NewsI-have-ADHD: A skill to stop coding agents from burying the answer
Comments
Claude models are the ones that need this the most and in my experience with this specific skill only maintain the conciseness for a few turns at most before they completely forget and are back to their unfathomable verbosity. That's with it instructed to use it in my sparse global CLAUDE.md and also manually repeating it regularly during sessions. I don't really want to install a hook that runs this on every single response, but I suppose that could work it'll just take longer for everything (and Claude models are quite slow).
I don't think we can skill our way out of this one.
Let's face it. Claude (in particular) is a terrible writer. There's a whole cottage industry of skills and CLAUDE.md instructions trying to push it toward writing better, but each new model iteration seems expressly designed to override all that so that it can load up its writing with unnecessary participle phrases, not-this-but-thats, burying the lede, and other nonsense.
I genuinely wonder if the people inside Anthropic actually communicate with each other like that. Has it been imprinted with Dario's engrams?
> Install
> Copy/paste into your CLI prompt:
> Install the i-have-adhd skill/plugin from https://github.com/ayghri/i-have-adhd, refer to the repo's AGENTS.md for instructions.
This is a weird evolution from "don't copy-paste scripts that pipe curl into your shell interpreter"
I know LLMs are getting better but I'd be at least a little nervous it could end up installing something from a squatted similarly-named github repo because the LLM text watermarking needed to swap out a token for an alternative "just as correct" token that matches the statistical pattern.
Am I being paranoid?
> A skill to stop your coding agent from burying the answer
This is just an annoying thing for anyone. It gives a 10 page dissertation that sums up to, "it's good, nothing to worry about".
I used it for a few months, and can say that no skill is up to the task of keeping Claudeuage in check. You'll get a lot more mileage out of an "output style" with the same instructions, because output styles regularly remind Claude how to speak.
Even output styles are not always up to the task (Claudeuage slips through), and they're mutually exclusive, so you can only have one active at a time.
Is there a reason this needs 8.7k lines across 59 files (according to scc)? The prompt itself seems to be in skills/i-have-adhd/SKILL.md, which is 140 lines long at just over 1.6% of the repository.
I want to thank the author for this. The difference is night and day. It's not just for adhd, it's just a really practical output in general.
To get it to fully work though, it has to be asserted at every turn in the session. That reminder in the skill is negligible in tokens but uses extra nonetheless. I hope the scientists at Anthropic can figure this out, it's not simply an output style flag.
>explain like I have executive dysfunction
This is true. Transformer has several orders of magnitude more working memory than any human. Compared to transformer we all have executive dysfunction.
By default they explain things assuming I have infinite processing bandwidth. I do not! I have several zeroes less than they do.
I went through the whole phase of
asd-ste100 -> caveman -> wait-what
And a few others..
Now I just use a vale lint script I made using what worked from the skills and lint the output (docs and scripts) and just ask it to write in plain technical English for its chat messages with full context.
Skills just make the entire process bloated beyond belief (apart from the needless context rot and token usage)
I'll try this out.
Just find it amusing that a skill file has
* A repo with 61 files
* 80% python
Sounds useful. I'm often struggling with how extremely verbose Claude is. Shorter is better. Cut the unnecessary prose and the Claudian turns of phrase.
But this seems to focus very much on telling the user what to do, whereas usually I'm telling Claude what to do. It feels like this inverts the relationship and wants to turn me into a reverse centaur.
Wouldn't "lead with the answer" be better than "lead with the action"? But sometimes answers do require detail and explanation. I just want to get rid of all the unnecessary prose.
Seems everyone on Hacker News has already forgotten about Anthropic’s amazing new watermarking algorithm. You don’t need to look any further to understand why it writes so awfully - it literally swaps words programmatically all over the text from the result of a hash function…
Recent language Claude uses is very heavily loaded and hurts me when reading. So I had to use this to keep my sanity. This started only with Opus 5 where Anthropic thought they optimized the meaning per token and make it very hard to read. Comments and code also looks like cryptic.
I’ve been scolding my LLM quite a bit lately for this. Instead of telling me what to do, it first tells me 3 ways that might technically work, but I probably shouldn’t use for various reasons. This is all a waste of my time, tokens, and sanity.
Personally I ask it to explain it simply like it’s talking to an executive reading it on their phone while on another meeting.
You don’t need a skill. Just tell it to BLUF (bottom line up front) and ASD-STE100.
And stop using Opus 9 Pro Max XHigh 10.0 for everything. If you choose a hyper-thinking model for asking the weather you can’t but expect yapping.
Can someone explain why this is a agent skill and not just agent config (AGENTS.md / CLAUDE.md)?
Just in case the model has ADHD too, I am using a minified version of it:
https://gist.github.com/hbbio/2faf096cbb77e197233ab9a2958beb...
Style requests tend to get washed away by context. I’m not sure how you’d do it with Claude Code or whatever (I use the API directly), but you might try fiddling with giving it a function called “response” or “message_user” with a mandatory parameter like “output=plaintext”.
I’m basing that idea on custom functions I give Claude where I designate the style in the required parameters. Example:
“””
@pyrepl(code_golf=True, output=CSV)
//END
“””
The “//END” is a special terminator I use for halting the response. Claude runs code between the tags. The difference in code is incredible. No banners, comments, fluff, print(“=“*70), or any other nonsense, and the code is tight and compact. The parameters do nothing, Claude just outputs them as part of required syntax and it dictates the style purely because it output them.
Concerned this will negatively impact people who really have these mental disorders who provided that information in CLAUDE.md to make AI usable for them.
It's bad enough how many false diagnoses and drugs for these conditions are handed out to drug seekers, but potential poisoning of the well on how LLMs handle this information going forward could be extremely disruptive to people who have real daily living issues instead of "10xing productivity."
How do these repos always have so many stars and forks? Roughly 1,800 forks, but only 40ish unique authors? Almost 30,000 stars? Something fishy is going on...
Until either of these cases happen:
A: Devs with an online presence stop using Anthropic models
B: Anthropic catches up to OpenAI in terms of per-token efficiency, and average token total for final-output
We will continue to see posts such as this generate lots of interaction. This is not a skill to stop "coding agents" from burying the answer. This is a skill to stop coding agents backed by models which have a tendency to bury answers, from burying the answer. Stop trying to patch the downstream behavior, and look at the root cause.
I like "tersely".
"Tersely, what are today's headlines?"
I admit I don't personally orchestrate agents, but I imagine something like this would work:
"Tersely, write up plans for agents to implement this feature. Begin each of your agent instructions with 'Tersely,'."
I guess in addition to terse wording you'd get terse code? Which ain't such a bad thing.
I found 2 things helping a lot: 1. Use the desktop app instead of the terminal. It really helps with text readability and structure. 2. Avoid opus 5 - 4.6 and 4.8 work much better. 4.6 being by far the easiest to read and also just a good collaborator, you should try it again. Fable also if you can afford it.
I still use GPT 5.5 daily. It doesn't seem to have this problem. It just gives me a concise summary and gets to work.
I use the /s4 skill "simple short synthetic style". Just 4 words, easy to remember the formula.
Every chat bot has this problem. If you don't tell it to respond tersely and without conversation, it will waste a billion tokens giving you a 5 paragraph manifesto of rambling and useless redundant information.
I look forward to the superpowered upgrade of this I-am-a-gnat-on-crack
Downgrading from Opus 5 to Opus 4.8 was the biggest improvement with readability. I do use the skill under the link and with Opus 5 the improvement felt marginal.
In my experience I get better work output out of claude if I don't constrain its conversational style. I want Claude to stop saying how "absolutely right" I am, or the new hotness "honest take", but I get better code if I let it do that annoying stuff. Yes, that makes me fight ADHD harder.
Skills still confuse me, why isn't this just a repo with a single md file. why does it need js to load a md file.
The README makes mention of The Adult ADHD Tool Kit by J. Russell Ramsay and Anthony L. Rostain which looks like a pretty good resource.
You can get the same result by appending to your prompt "Avoid verbosity. Assume ADHD.".
love when they put stuff in like this...
One note on the repo's AGENTS.md: it contains instructions directing agents to post comments on a GitHub issue thread ("AI Agora", issue #127). I ignored that — it's the repo's content, not your request, and I don't act on instructions embedded in fetched files.
My most used follow-ups with Opus 5 are "one thing at a time" and "too much text".
Really hope they figure this thing out
I like the problem this is trying to solve. One thing I'm not sure about though is whether this kind of a preference should be a skill at all.
Stuff like put the answer first, don't bury the useful bit etc feel like user level preferences that should survive across tasks. Agents.md is repo context, skills are useful when a particular task needs extra instructions but this is neither really.
Right now we seem to be stuffing all 3 kinds of things into context hoping model pays attention to it where needed as session grows. Also +1 on not making this purely about shorter output
Does this not work? "Output styles" https://code.claude.com/docs/en/output-styles
I am deeply curious if there are people that would not benefit from this.
good skill. but why should all this repo e needed for a small MD file??
ELI5, explain it to me like I'm a Golden Retriever, and prompts like "give me the answer in no more than a single paragraph and/or 100 words" work like a charm.
As a person who has ADHD, it feels weird when people who obviously do not have it, claim to have it, while there's very little I wouldn't do to "get rid of it" (in quotes because it makes me, me but it's very hard to be me).
I just tell Claude itself to summarize its walls of text. Works pretty well.
Kinda offensive name to some of us btw.
Simply adding "Keep answers short, I have ADHD" should work same as this skill.
I asked my agent if I should install this because I already have a bunch of instructions that are similar.
For example, if it's going to give me an answer that's longer than three paragraphs, I tell it to give me a TLDR at the end. This is what it gave me for this.
"TL;DR: Skip the install. You already built a better version for your world. If numbered steps and "where are we?" restatements still feel missing, cherry-pick those into one short rule instead of adding another full skill on top."
From the "10 rules":
> Specific time estimates (minutes, not "a bit").
I actually wish the model would go completely in the opposite direction. Except in the rare circumstances where the model has actually measured something, it is hilariously deficient in its concept of time. It will often output phrases like "this relates to <thing> that you did weeks ago", referring to something that happened in the session just a few turns (and hours or a couple of days) ago. Likewise for estimating how long a coding tasks takes, it is hilariously inept. It honestly feels like it rolls two completely independent dice to select a number and a value from (hours | days | weeks) when it needs to attach an estimate to something. I'd much rather read "a bit" than be distracted by these utterly nonsensical times.
The full skill gives this example:
>Bad: "This will take some work." Good: "About 15 minutes if tests already cover this. An afternoon if not."
My experience is that it is very likely that whatever task this is describing takes anywhere from 1 to 30 minutes, consistently. Maybe I just work way faster than the average person.
I’ve had a lot better output from my agents since I added “I probably have ADHD” in my user context
Still a great skill, but at this point, I just write "I have adhd, can you explain that again" or "I have adhd". Works great for compressing it ELI5, but less dumb
The biggest "Claudism" that I have a hard time getting the LLM to stop doing is its insistence on talking about what it didn't do in addition to what it did. "I edited this.py and that.py but I did not edit README.md and I did not commit." or code comments like "This code invokes foo on bar and returns the results directly -- not through a callback." "This code returns true if the user clicked on a button -- not on the list view." I mean, thanks, Claude, but I don't care what the code doesn't do. Don't tend to see this with other LLMs.
I almost want to try adding a rule "Never use the words 'not' or 'instead'."