Hacker News

Show HN: Deslop skill based on corpus-linguistics and psycholinguistics research

16 points by dannypovolotski ago | 5 comments
I vibed this deslop skill because I was getting increasingly annoyed that deslop skills for prose relied either on fitting prose into an overly restrictive format (simple-english and iso-24495) or telling AI to rewrite something without providing any tools in the skill to measure the outcome (which more often than not resulted in one slop being replaced with another).

This skill is based on published measurements of how LLM text differs from human text: Reinhart et al.’s comparison of Biber’s grammatical features in human and model writing, Terčon & Dobrovoljc’s survey of 44 studies, Clark, Martin & Mahowald’s work on why AI text is hard to read, and Graphite’s “AI Tells” study, whose released data includes tell lists for nine current models. It’s benchmarked against three corpora in which humans and models answered the same prompts: 9,984 web-article topics from Graphite, 24,322 questions from HC3 and 3,027 advice prompts from ELEPHANT.

It also ships with a scanner that runs before and after each revision. It shows which problems the revision fixed or newly caused, and it catches dropped facts and any tells from the reviser’s own model family.

hereme888 |next [-]

Looks promising. I've been using a modified version of simple-english with my agents, and it's only sometimes better than not using it.

Could you add a license file?

dannypovolotski |root |parent [-]

MIT license added. I've been using Simple English with mixed results. The biggest annoyance with it is that sometimes it ends up actually exaggerating some Claudisms I really hate, since it trends towards shorter sentences. Let me know if this approach feels effective to you after you try it!

hereme888 |root |parent [-]

It's working well!

I decided to combine multiple writing skills into one codebase, with yours as the default. Here's an example of what I consider significant improvement to one of my readme files: https://github.com/sm18lr88/VS_Code_Check/commit/1dfa3d31126....

I gave you access to my version of the skill, which I'm keeping private until I know it works well and makes sense.

fanquiz |next |previous [-]

[flagged]

pentaphobe |previous [-]

I like the definition of slop as "when we put less effort into producing something than we ask the consumer to spend understanding it." [^1]

rather than some amorphous "made by an LLM and I can tell"

That'd make "deslop" more of a "slop camouflage" since it just makes it less overt when a human has vibed the content instead of putting in any modicum of effort. :)

Still - impressive example in the readme, if we can't prevent slop then at least can reduce how much reviewer lifetime is wasted

[^1]: https://www.slopmanifesto.com/

dannypovolotski |root |parent [-]

I agree! I personally either heavily edit or rewrite consequential communication to others if it includes anything that was written by AI.

Often if I want to get over writers block, I'd describe a messy idea to an LLM and limit it to giving me THOUGHT_A_GOES_HERE and help me structure my writing, but then I actually write what I was going to.

This is more aimed at making the way an agent talks to ME more tolerable, or at least understandable.

I do, however, think that it goes beyond just "camouflage" because it addresses specific readability and understanding gaps that make the output meaningfully more useful to me when I read AI output and save time.