✎ Edit this page

Tutorial 3: Write your first skill

← Learn by building

You will: write a skill for one Buddhist task, test it by hand, share it with the community, and (if the task has a benchmark) have it measured. You need: a text editor and a GitHub account. An AI assistant that reads skills (for example Claude) helps you test. About 45 minutes.

What a skill is

A skill is a folder with a file called SKILL.md. The file has a short header and then written instructions: how to do the task, the rules, the output format and an example. An AI assistant reads it when the task comes up. A good skill makes a model do the task better, or makes a cheaper model do it as well as an expensive one.

One skill is for one task in one language. Its name is <task>-<language>, for example toc-extraction-tibetan or translation-pali-to-english. Task names are in the task list.

1. Copy the template

Copy templates/skill/SKILL.md from buddhist-skills into a new folder named after your task and language. Fill in:

2. Write the instructions

Write as if you explain the job to a careful new assistant who knows the language but not your project:

  1. Steps: what to do first, next and last.
  2. Rules: the things your team checks by hand. Which glossary, which transliteration, what never to change.
  3. Output format: exactly what the answer looks like. If the task has a benchmark, use the format in its TASK.md.
  4. Example: one short real input and the correct output.

Keep it short. Long skills cost more tokens on every call, and the leaderboard shows cost.

3. Test it by hand

Give your skill and three or four real inputs to your assistant. In Claude, add the folder as a skill, or paste SKILL.md into the chat and then the input. Check the answers yourself. Fix the instructions where the model goes wrong, and raise the version.

For a task with a benchmark, you can test on the real items: put the text of your skill into PROMPT in the script from Tutorial 2 and score it.

4. Share it

Open the Share your files form and paste a link to your folder (or attach the file). A maintainer adds it to the prompt library with your name in authors. You do not need to know git, but you need a free GitHub account. The form is public.

5. Have it measured

If the task has a benchmark, a maintainer runs your skill on it with the standard models, or you submit a run yourself (SUBMIT.md). The run gets a provenance record, so anyone can see exactly which version of your skill made the score. If your skill becomes the Top performance or Best value for its task, it is released in buddhist-skills and we announce it, crediting you.

If the task has a benchmark, a maintainer also adds your skill as a variant in the task’s workshop repo. If the task has no benchmark yet, your skill stays in the prompt library until there is one. You can help build it: Tutorial 4.

Good to know

✎ Edit this page · History · Discuss