Architecture
The flow
workflows → tasks → benchmarks → results → leaderboards → best skills
(knowledge base) (buddhist-benchmarks, buddhist-bench) (buddhist-skills)
↑
workshop repos (one per skill, e.g. toc-extraction-tibetan): variants, runs, reports
- A workflow page describes how a real project works. Each step maps to a task.
- A task has a definition page here and an entry in buddhist-benchmarks
tasks/index.json. - A benchmark (test data + scorer) measures the task for one language.
- Skills and models are run on the benchmark. Each run adds a
result.json. - Scripts build the leaderboards from the result files. Nothing on a leaderboard is typed by hand.
- The best skill per task and language is released to buddhist-skills, where people install it.
Repos
All repos are in the BuddhistAI GitHub org, public, MIT licensed. Each has a README.md and a CLAUDE.md / AGENTS.md that starts with the same “Start here” block (read this handbook, check the Development board, the core rules) followed by the repo’s own rules. The main repos also have a scripts/check.py that CI runs. The homepage and toc-extraction-tibetan have no CI.
| Repo | Role | Site |
|---|---|---|
| .github | Org profile (profile/README.md, banners in profile/assets/), org-wide CODE_OF_CONDUCT.md. |
github.com/BuddhistAI |
| template-skill-repo, template-benchmark | GitHub template repos for new skill repos and new benchmark repos (how-to). Created 2026-11-18. | — |
| buddhistai.github.io | The homepage. One self-contained index.html (inline CSS and JS). Scores on it are copied by hand from the leaderboard. |
https://buddhistai.github.io/ |
| design | Design system: tokens.json (source of truth), generated buddhistai.css, tailwind.preset.js, tokens.js, logos, banners, VOICE.md. |
https://buddhistai.github.io/design/ |
| knowledge-base | This site: task definitions, workflows, experiments, Papers (community papers with Markdown full text), the Directory, the Languages pages (AI resources per Buddhist language and a country overview), and this handbook. Jekyll on GitHub Pages. Was called wiki; the old research repo was merged in. |
https://buddhistai.github.io/knowledge-base/ |
| buddhist-benchmarks | Hub for task benchmarks: tasks/index.json, per task TASK.md, scorer, data, results; models.json (prices); external.json (other groups’ leaderboards); STATE.md; SUBMIT.md. Builds the leaderboard site. |
https://buddhistai.github.io/buddhist-benchmarks/ |
| buddhist-bench | BuddhistBench: a Q&A benchmark of Buddhist knowledge (tracks by tradition, text and exam; language as an axis). Also the OpenRouter runner and the intervention study. Version 0.4.0, examples plus one imported set. | Shown on the leaderboard site |
| buddhist-skills | The shelf: only released (promoted) skills in skills/ (per task and language, the top performance and the best value skill), the router, shared glossaries in shared/. Builds the Claude plugin marketplace. community/ and drafts/ are one-line pointers since 2026-11-18. |
Installed as a plugin |
| buddhist-prompts | Shared and not measured: personal AI files (CLAUDE.md), project instructions, style guides, glossaries, prompts, and skills or workflows with no score yet. <topic>-<language>/<author>-<name>/ with CARD.md (kind). Holds the Share form. Created 2026-11-18. |
GitHub only |
| toc-extraction-tibetan | The first workshop repo: every variant of the Tibetan TOC skill, the runs, and REPORT.md. Pins the benchmark; holds no test data. |
— |
| research | Archived. Now the Experiments section of this site. | — |
Old names still redirect: BuddhistAI/skills → buddhist-skills, BuddhistAI/wiki → knowledge-base. Always use the new names.
Hugging Face: org BuddhistAI. Data sets there are copies made by CI (below). Never edit them on Hugging Face.
Target layout (decided, not built yet)
The team decided on 2026-11-17 that each benchmark gets its own repo with maturity information, and buddhist-benchmarks becomes a list of all benchmarks per task. Today the only task benchmark (toc-extraction/bo) still lives inside buddhist-benchmarks. Likewise each task gets a workshop repo with all submitted versions, and buddhist-skills holds per task both the Top performance and the Best value skill (today it holds one released skill per task and language). See Status.
Sites
| Site | Built from | How it deploys |
|---|---|---|
| https://buddhistai.github.io/ | buddhistai.github.io index.html |
GitHub Pages from main, root. No build step. |
| https://buddhistai.github.io/buddhist-benchmarks/ | buddhist-benchmarks scripts/build_leaderboards.py --site + site/ (CSS, JS) |
leaderboards workflow (.github/workflows/pages.yml) on push to main and daily at 03:17 UTC. Pages source: GitHub Actions. |
| https://buddhistai.github.io/knowledge-base/ | knowledge-base Markdown, _layouts/default.html, assets/ |
GitHub Pages from main, root (Jekyll, plugins in _config.yml). |
| https://buddhistai.github.io/design/ | design index.html |
GitHub Pages from main. |
buddhistai.org |
— | Domain bought by NT on Cloudflare (2026-11-17). DNS is not set up yet. |
GitHub Projects boards
All boards in the BuddhistAI org are public (NT, 2026-11-18): anyone can see them, only members can edit. Keep nothing on a board that cannot be public. The Skills, Benchmarks and Prompt library boards are for contributors; the Development board is for team work. Columns, fields and automations are set only in the GitHub web UI (Settings of each project).
| Board | Columns | Filled by |
|---|---|---|
| Skills board | Requested → Shared → In progress → Being measured → Released | Auto-adds open buddhist-skills issues labeled request or community-share. Share issues now open in buddhist-prompts and go to the Prompt library board (below). |
| Benchmarks board | Proposed → Needs test data → In progress → Being reviewed → Live | Auto-adds open buddhist-benchmarks issues labeled benchmark. |
| Prompt library board | New → Reviewing → Listed | Auto-adds every open buddhist-prompts issue (Share your files). Created 2026-11-18. |
| Development board, for team members | To do → In progress → In review → Done (to add by NT: a Design review status after To do, see UX reviews); fields Owner and Area; views Board and Table | Team tasks, added by hand. Public since 2026-11-18. |
Issue forms and labels
| Repo | Form | Label | What a maintainer does |
|---|---|---|---|
| buddhist-skills | Request an AI solution for a task | request |
Find or define the task; add a task page here if needed. |
| buddhist-prompts | Share your files (a link or uploaded files: a skill or any file given to AI) | community-share |
Turn it into a prompt library entry and credit the author (how-to). The old buddhist-skills form is a stub that links here. |
| buddhist-skills | Volunteer as a language reader | language-reader |
Match them with data that needs a reader (see Status). |
| buddhist-skills | Report a problem | bug |
Fix or route. |
| buddhist-benchmarks | Propose a benchmark | benchmark |
Discuss, then add the task and data (how-to). |
| buddhist-benchmarks | Review a benchmark as an expert | expert-review |
Agree on the scope, share the items, record the review (how-to). |
| buddhist-benchmarks | Donate compute or API credits | compute |
NT replies; set up the key as a secret; credit the donor (how-to). |
| knowledge-base | Add to the Directory | directory |
Verify the link, add a YAML entry (with access and collaborative for datasets). |
| knowledge-base | Update a language page | (edit link) | Edit languages/<id>.md; each fact needs a source. |
| knowledge-base | Share a workflow | workflow |
Write a workflow page, credit the owner. |
| knowledge-base | Suggest a task | task |
Write a task page and a planned entry in tasks/index.json. |
| knowledge-base | Idea (experiment) | idea |
Discuss; if it goes ahead, write PLAN.md. |
| knowledge-base | UX feedback | ux-feedback |
Sort weekly: a board card for a problem, a persona’s “Known pain points” for a fact about users (UX reviews). |
| knowledge-base | Re-run, To-do | re-run, todo |
Track work. |
| buddhist-bench | Propose items, Propose a track, Dispute an answer (Markdown templates) | — | Follow buddhist-bench CONTRIBUTING.md: real items need human authors and two reviewers. |
Pull request templates: buddhist-skills, knowledge-base, buddhist-benchmarks, buddhistai.github.io and design. Each has an “EGM review” section (UX reviews).
Community: GitHub Discussions
NT chose GitHub Discussions as the community home (2026-11-17): https://github.com/orgs/BuddhistAI/discussions. Start with the welcome post (pinned).
Categories: Announcements (maintainers only), Introductions, Find the best AI (Q&A), Requests and ideas, Skills, Models and evaluations, Benchmarks and test data, Tasks and workflows, Languages (put the language in the title), Arena and judging, Website and development, Polls.
- Point people to the matching category when they ask “where do I ask?”. Bugs and concrete requests still go to the issue forms above.
- Maintainers post leaderboard news and new leaders in Announcements.
- To get notified: watch the BuddhistAI/.github repo, choose Custom, then Discussions.
- A chat channel (Discord or WhatsApp) may come later. Not decided.
CI workflows
| Repo | Workflow | When | What |
|---|---|---|---|
| design, knowledge-base, buddhist-benchmarks, buddhist-bench, buddhist-skills | check |
Push to main, pull requests |
Runs check.py. A failing check means a generated file is stale or a rule is broken. |
| buddhist-benchmarks | leaderboards (pages.yml) |
Push to main, daily 03:17 UTC, manual |
Rebuilds LEADERBOARD.md files, the README table and STATE.md and commits them ([skip ci]); creates missing data release tags <task>/<lang>@X.Y.Z; builds and deploys the site (reads buddhist-bench results and live top rows of external.json boards). |
| buddhist-benchmarks | hf-sync |
After leaderboards, on tag push, manual |
Publishes each data release to Hugging Face and the results table to BuddhistAI/buddhist-benchmarks-results. |
| buddhist-bench | hf-sync |
VERSION changes on main, v* tag, manual |
Checks, tags vX.Y.Z, publishes public splits to BuddhistAI/buddhist-bench. |
| buddhist-bench | daily |
Daily 03:00 UTC | Intervention-study runs on OpenRouter. Skips cleanly without keys. |
| buddhist-bench | import-nak-tham |
Manual | Imports the Thai exam set from Hugging Face (Actions can reach Hugging Face when a container cannot). |
| knowledge-base, buddhistai.github.io, buddhist-benchmarks, buddhist-skills, design | egm |
Pull requests | Calls egm-gate (free: a pull request that changes pages users see needs an EGM design or “Small change”) and, with the label egm-review, egm-goldfish (one Goldfish review round, posted as a comment; needs ANTHROPIC_API_KEY). See UX reviews. |
| knowledge-base | egm-gate, egm-goldfish |
Called by egm |
Reusable workflows used by every repo’s egm workflow. |
| knowledge-base | ux-walkthrough |
1st of each month 05:41 UTC, manual | Claude reads the live sites as each persona and opens a ux-walkthrough issue. Needs ANTHROPIC_API_KEY. |
| knowledge-base | agent-notes |
Mondays 04:23 UTC, manual | Opens a pull request adding the standard CLAUDE.md and AGENTS.md (from scripts/agent-notes/CLAUDE.md) to every org repo that has none. Needs ORG_REPOS_TOKEN. |
Tags pushed by GITHUB_TOKEN do not start other workflows. That is why hf-sync runs on workflow_run after leaderboards.
GitHub to Hugging Face
GitHub is the source of truth. Hugging Face holds copies made by CI.
- A data release tag
<task>/<lang>@X.Y.Zin buddhist-benchmarks becomes the datasetBuddhistAI/<task>-<language name>(for example BuddhistAI/toc-extraction-tibetan) with Hugging Face tagvX.Y.Z.mainthere is the newest release. It ships the data folder,TASK.md, the CHANGELOG, the scorer and the license. - All result files become one table:
BuddhistAI/buddhist-benchmarks-results. - buddhist-bench publishes
BuddhistAI/buddhist-bench(publicdevandexamplesplits only; nevertest, never held-out exam papers). - Details: buddhist-benchmarks docs/huggingface.md and buddhist-bench docs/huggingface.md.
- Live since 2026-11-17. Both workflows do a dry run and skip publishing when
HF_TOKENis missing.
Secrets
Never commit a key. Repo or org secrets are set by NT in GitHub Settings → Secrets and variables → Actions.
| Secret | Used by | Status (2026-11-17) |
|---|---|---|
ORG_REPOS_TOKEN |
agent-notes in knowledge-base (opens PRs in other org repos); set it as a repo secret of BuddhistAI/knowledge-base |
Not set. Fine-grained token, resource owner BuddhistAI, all repositories, permissions Contents: read and write, Pull requests: read and write. Without it the workflow only lists the repos that need notes. |
HF_TOKEN |
hf-sync in buddhist-benchmarks and buddhist-bench |
Set. Write token of the BuddhistAI Hugging Face org. |
OPENROUTER_API_KEY |
buddhist-bench runner and daily workflow |
Not set. Needed for live model runs. Dave offered sponsor keys. |
BRAVE_API_KEY |
buddhist-bench logged-search mode |
Not set. |
ANTHROPIC_API_KEY |
Workshop runs on your own machine (toc-extraction-tibetan bench_skills.py): local only. As an org secret for the egm-goldfish and ux-walkthrough workflows (UX reviews). |
Not set as a secret (2026-11-18). Without it, those workflows skip. Put a monthly spending limit on the key. |
Automatic scoring of submissions (planned) will need an inference key as a secret too. Use a sponsor key, and put a spending limit on it.
Permissions
- NT has given standing permission to merge PRs, push to
main, and create repos in the org without asking first. - CODEOWNERS: NT reviews everything until language and workflow owners exist. A hosted workflow’s owner is added on the line for their page.
- Access model (written 2026-11-17, see Governance): anyone contributes by fork and pull request or the forms. A partner organization gets a team
@BuddhistAI/<org>-teamwith Write only on its own task repos (optional Triage on buddhist-skills and buddhist-benchmarks, and CODEOWNERS lines for its language). Only a Maintainers team has write access to the core repos (buddhist-skills, buddhist-benchmarks, buddhistai.github.io, design, knowledge-base, .github). Releases into buddhist-skills happen only through the leaderboard. - Teams (2026-11-17):
Maintainers(@BuddhistAI/maintainers) andWeBuddhist(@BuddhistAI/webuddhist) exist; NT is in both; no repo permissions given to either yet. - Not set up yet (NT decided on 2026-11-17 to document now and set up when the first partner team joins): Maintainers access on the core repos, member base permission check (must be Read), and a branch rule on
mainof each core repo requiring a pull request with one code-owner review. Today nobody outside NT and Claude sessions has write access, and they push tomaindirectly. - Creating a new repo needs an org owner (NT) in the GitHub UI. Always start from a template (how-to).

BuddhistAI Knowledge base