What languages are agent skills written in?

📝 Discussion Summary (Click to expand)

Underrepresentation of non‑English code in training data – Researchers may miss many Chinese repositories, leading to biased skill estimates.

“I wonder if the Chinese skills are undercounted somewhat‑ there could be more repositories the researchers didn’t have access to.” – JSR_FDED
Effect of language on model reasoning and output quality – The language a model is prompted in can influence whether it “thinks” in English and how well it performs.

“Would love to know how changing the language impacts the output quality. Do the models using these non‑English skills still reason in English?” – njdancer
Linguistic relativity and expressive power of different languages – Certain ideas may be clearer or easier to express in some languages, affecting what models can capture.

“Do different languages make it easier to express certain ideas more clearly? Considering how linguistic relativity might apply here.” – njdancer

🚀 Project Ideas

A programmable toolkit that measures how changing the input language affects LLM outputs, exposing reasoning consistency and cross‑lingual performance gaps.
Core value proposition: gives researchers a quick, reproducible way to quantify language‑specific model behavior and uncover hidden biases.

Key	Value
Target Audience	ML researchers, LLM engineers, AI auditors
Core Feature	Automated prompts in multiple languages with metrics for equivariance, accuracy drift, and chain‑of‑thought consistency
Tech Stack	Python, HuggingFace Transformers, spaCy, Pytest, optional FastAPI for serving
Difficulty	Medium
Monetization	Hobby

HN users expressed curiosity: "Would love to know how changing the language impacts the output quality." (njdancer)
Enables concrete experiments on linguistic relativity hypotheses and can spark discussions about model generalization across languages.

Scrapes public code and text repositories to estimate the proportion of non‑English content, highlighting potential undercounts of skills like Chinese.
Core value proposition: provides concrete data to validate or refute claims about language bias in training corpora.

Key	Value
Target Audience	AI researchers, data curators, model auditors
Core Feature	Language detection pipeline that reports per‑language token counts and trends over time for sources like GitHub, Common Crawl
Tech Stack	Python, GitHub API, fastText or langdetect, Pandas, Docker
Difficulty	High
Monetization	Hobby

Directly addresses JSR_FDED's concern: "I wonder if the Chinese skills are undercounted somewhat..."
The resulting statistics can fuel HN debates about data provenance and guide more balanced data collection efforts.

Interactive web app where users submit a prompt in several languages and view side‑by‑side LLM responses, with highlighted differences and optional chain‑of‑thought toggling.
Core value proposition: makes it easy to experience and measure how language shapes model reasoning, turning abstract linguistic theories into tangible observations.

Key	Value
Target Audience	Developers, linguists, product managers, educators
Core Feature	Multi‑language prompt interface, response diff visualization, exportable analytics, optional reasoning‑mode toggle
Tech Stack	React, Node.js/Express, Redis for caching, LLM API (Open‑source or commercial), Tailwind CSS
Difficulty	Medium
Monetization	Revenue-ready: subscription tier for team usage ($15/mo per seat)

Mirrors njdancer's request: "Would love to know how changing the language impacts the output quality."
Encourages practical discussion on linguistic relativity and can be used in education or product design to test multilingual features.