Hacker News Reader

Comment by sleazebreeze | Hacker News Reader

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.

View on HN

Claude Code has output styles; one of the predefined styles is Concise [1].

    Concise: Claude leads with the result, skips preamble and narration, and
    keeps responses short by default, while doing the engineering work as
    thoroughly as in the Default style. When you ask for an explanation or
    more detail, Claude answers in full. Claude always keeps the complete
    content of error reports, security warnings, and confirmations for
    destructive actions. Requires Claude Code v2.1.237 or later.

[1]: https://code.claude.com/docs/en/output-styles

Output styles are just more context (like CLAUDE.md) with a tiny bit more weight. They very much don't force Claude to speak succinctly, no matter what you put in them.

I got so angry at Fable's verbosity that I had a non-Anthropic model vibe-code a dedicated English linter. One that threw spaghetti at the wall in the form of combined Flesch-Kincaid, Gunning Fog, SMOG, Coleman-Liau, ARI and a whole pile of regexes and NLP as a fast Simple Technical English Gate. It also incorporates nicely into other agent projects, like for comment and README linting.

So if it gets bad I simply tell it "I ain't reading all that, feed it through the STE Gate" and it will tame the results. I haven't bothered to set it up as a hook yet.

Wanna share this?

In a way I already did, my comment is practically the prompt itself.

This sounds amazing. Please share!

OTOH, Claude also forgot its own system prompt as turn passes. Claude is almost jailbreaked once you virtually gain its confidence and will answer prompts it would never have answered is you asked right at the start of the conversation.

It's weird to be able to use social engineering against a program.

> for a few turns at most before they completely forget

Which sounds more like Claude has ADHD than the user does.

Same experience — it's not that the instruction stops working, it just gets drowned out as context grows. Re-stating a short version periodically works better for me than a one-shot instruction up front.

GLM (both mainline and flash varieties), being trained to speak a very similar Claudeish, is also atrocious about this. My AGENTS.md demands all prose be in STE100, which is trying to solve a similar problem as the skill linked in OP. Getting GLM to STFU by AGENTS instruction is nigh impossible. I use a plugin for OMP/Pi to periodically inject system prompt reminders of output style, and it still largely ignores the request, dumping novels of jargon-dense prose every turn.

DeepSeek v4 Flash isn’t much better (unsurprising- it’s an extremely stubborn model).

Weirdly, GPT Luna excels at following this type of instruction from AGENTS.md, and never forgetting it, even 400k+ tokens into the context window.

yup.

GPT Luna tends to keep things objective. Muse Spark 1.3 is also one of the better models in this aspect, for me.

I have such hook, and it works well. It increases time, but it’s negligible if agent works for 10+ mins before the final output. The hook is only on stop event, so agent is not interrupted during the session. I didn’t have luck in putting it to skill / md files, as agent just ignores it after a couple of iterations. The solution is not perfect, sometimes it compresses too much, and I have to reread both versions, but at least I don’t have to read load bearing stuff anymore

Reply on news.ycombinator.com