Table of contents
MeasuredFinds the sa-bcad outline headings in a commentary, with their level and position.
We test AI models and skills on real Buddhist work and share all the results for free. For each task, see the top performance (the highest score) and the best value (a high score for much less money). If you make AI models or skills, send us yours and we will test them too.
Skills, models, workflows or test data. Paste a link and we score it on the open benchmarks for free. Lead a task and we announce it.
No repo? Share your files: a skill, project instructions, a Claude file, a style guide or a prompt. Needs a free GitHub account (make one). The form is public.
Grouped by the kind of work. Each task shows two picks: the highest score, and the best score at a fifth of the cost or less.
One plugin per language holds the top skill for every task. Pick the best-value version if cost matters more than the last few points.
/plugin marketplace add BuddhistAI/buddhist-skills/plugin install buddhist-skills-tibetan@buddhistaitoc-extraction-tibetan.zipFinds the sa-bcad outline headings in a commentary, with their level and position.
Fix OCR and input errors and normalize a text without changing its content.
Split a text into sentences or translation units. In progress with the Tibetan team.
People, places, texts and other named entities.
Find quotations and link them to their source texts.
Link each root-text passage to the commentary that explains it.
Propose glossary terms and renderings from a source and its translation.
Draft translation. Scored by side-by-side reader votes as well as automatic metrics.
Check a translation against its source, glossary and style guide.
Thangka and iconography description, image prompts that respect tradition, art cataloguing.
CLAUDE.md files and project instructions for practitioners, teachers and dharma centers.
Sharing needs a free GitHub account (make one). The form is public.
All tasks on one page. Models are ranked on every benchmark with no skill, including Q&A benchmarks. Skills are ranked on top of each model; the top skill is the one in buddhist-skills, and a challenger that beats it replaces it.
| # | Skill | Model | Score | Gain | Cost / 100 pages |
|---|
| # | Model | TOC, Tibetan no skill | With best skill | Q&A, Thai Dhamma exams | Cost / 100 pages |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5Top performance | 0.951 | 0.962 | – | ≈$0.76 |
| 2 | Claude Haiku 5.5Best value | 0.741 | 0.824 | – | ≈$0.034 |
| – | GPT, Gemini, Qwen, DeepSeek, Llama… | not run yet | – | – | – |
Some work has no single right answer. Translators and teachers vote blind between two answers, from a model on its own or a model with a skill, and the votes become a ranking for each task and language.
| # | Entry | Rating | Votes |
|---|---|---|---|
| 1 | Skill A + model X | – | 0 |
| 2 | Skill B + model Y | – | 0 |
| 3 | Model X, no skill | – | 0 |
Tell us what you need to do. Every request goes on the public Skills board. People can vote for it, and developers can pick it up. The most wanted tasks get a benchmark first.
Requests and their votes are on the public Skills board. Add yours, or vote for the tasks you need.
Open and collaborative. Each benchmark lives in its own repo with a maturity level, so you know how much to trust a score. Groups may publish competing sets; new skills are scored on all of them.
| Benchmark | Ranks | Language | Items | Maturity | Maintainer |
|---|---|---|---|---|---|
| TOC extraction 2.0.0 | Skills | Tibetan | 85 | Draft, answer key needs a reader check | BuddhistAI |
| Thai Nak Tham exams | Models | Thai | 4,075 | Reviewed, public, may be memorized | Imported |
| Tibetan OCR | OCR engines | Tibetan | – | Stable, run by BDRC | BDRC |
| BuddhistBench Q&A, by tradition | Models | Many | – | Example items, 15 tracks open | BuddhistAI |
Have exam papers or a test set your team already uses? Paste its repo to publish it as a benchmark.
Short guides that chain tasks together, written with the teams who do this work. Green steps have a measured skill today.
From an OCR file to a structured, segmented source with its outline.
Collect renderings from past translations into one reviewed glossary.
Check a draft against source, glossary and style guide before an editor sees it.
Looking for other Buddhist AI tools or data? The Directory lists Buddhist AI projects and datasets from all traditions and languages. Add yours. To choose an AI for translating into a Buddhist language, see Languages: data, AI projects and translation quality for each language, and an overview of Buddhist countries.
Skills and benchmarks here are open source and built together. Anyone can add a skill, a model, a test set, or a fix to any of them. Paste a link to get started; evaluation is free, paid by a sponsor.
Today your link opens a GitHub issue and a maintainer runs the evaluation. Automatic checks are coming.
The prompt library holds files people give to AI: your CLAUDE.md, project instructions, a style guide, a glossary or a prompt. Drag your files into a short form. You need a free GitHub account, and the form is public. Your name and picture are optional.
Share your filesMade something? Show it in the community (Skills category).
Every skill's task repo keeps all its variants. Fork the top one, change it, paste your fork. If it scores higher, it becomes the main skill.
Add the edge cases your team runs into, fix a wrong answer, or publish your own set. Competing sets from different groups are all kept.
Read two answers and pick the better one. Votes from translators and teachers rank the open-ended tasks.
Know the material well? Check a benchmark's questions and answers in your field. You are named as reviewer, and the benchmark moves up to Reviewed. Offer a review
Every score costs money to make. Credits pay for evaluations of everyone's submissions and never change a score or a ranking. Offer credits
Short tutorials: run the scorer, evaluate your first model on a Buddhist task, write your first skill, add a benchmark. Start learning
The homepage reads from five kinds of sources. Each surface above is built from them, so nothing on the page is typed in by hand.