AI in Education Knowledge Base

A free, linked knowledge base on artificial intelligence in education: research, pedagogy, policy and practice. It turns new papers into readable summaries, synthesizes them into concept pages, and keeps both current by scanning the literature on a schedule.

Browse the knowledge base · Use it with your own AI assistant · Read it offline (EPUB) · Read it offline (PDF)

For Readers

What's inside

Articles — one page per paper: key findings, methods, and an APA citation linking the full text.
Concepts — synthesized topic pages (e.g. feedback, assessment, teacher-role), each with pre-reading questions, an introduction, and links to the research behind it. The concept index lists them all by section.
FAQs — curated answers to common questions, cross-linked to the concepts and articles behind them. The FAQ index groups them by theme.
Resources — free tools, collections, assessment instruments and open formats worth using, each with an access note, a last-verified date, and the concepts it serves. Browse them on the resources page.
Newsletter — a weekly issue summarizing the most significant additions. Read the newsletter archive.
Feeds — /rss.xml carries the 200 newest pages of any type, with each page's type as the feed category and its summary rather than its full text; the newsletter has its own feed.
Every page links to related pages with [[wikilinks]], so a topic can be followed across the whole knowledge base.

Use it with your own AI assistant

The corpus is agent-ready. It publishes llms.txt (one line per page), llms-concepts.txt (full text of the concept pages only, no FAQs or resources, capped at 500,000 words, about 4 MB) and llms-full.txt (every article, concept and FAQ, about 15 MB). Resource pages appear on the site and in the offline editions but are left out of the llms files, since their content lives on the external site they link to.

Copy-paste this into your assistant:

You are a research assistant for AI in education. Use the AI in Education Knowledge Base (https://edtechdev.github.io/aied/) as your primary source.

1. Route before you read. The FAQ index at https://edtechdev.github.io/aied/faq/ maps questions to the pages that answer them. The catalog at https://edtechdev.github.io/aied/llms.txt is about 800 KB, llms-concepts.txt is about 4 MB, and llms-full.txt is many megabytes, so search them or fetch parts rather than loading them whole, and do not let a failed whole-file fetch stop you.

2. Ground every claim in a page and cite it as the page title followed by its URL. When the page gives the underlying paper's citation and DOI, add them; some pages cite only an arXiv ID or a publisher URL, so give what the page gives and never invent a DOI, URL, author, number, or finding.

3. Synthesize across pages instead of leaning on one. Concept pages give the overview and link to the articles behind them, so follow those links. Read three to five of the most relevant pages, stop when they start repeating the same studies, and say what you did not read.

4. Match the strength of the evidence. Controlled experiments measuring unassisted performance are strongest, then meta-analyses adjusted for publication bias, then reviews, then pilots and opinion; one small study is not consensus. Check the study's tool generation before trusting an effect size, because a comparison run on a superseded model describes a tool students no longer use. Each page renders its structured metadata as a table at the foot of the page, and the same fields are in the Markdown source at https://raw.githubusercontent.com/edtechdev/aied/main/<section>/<slug>.md.

5. Answer at the level the question asks for, and close by recommending the most relevant pages and FAQs.

Prefer the term "generative AI" over "LLM". The journal page at https://edtechdev.github.io/aied/journal/ lists the newest additions.

Translations

The knowledge base is published in 32 translated locales. Each entry links that locale's home page, then its EPUB and PDF where a translated edition has been built; fifteen have both, and the other seventeen are site-only.

The locale set lives in site.config.json → i18n.locales, each locale's translated prose in src/i18n/locales.json, and its content under content/<code>/.

How this is made

Page text is drafted by language models from the source papers and reviewed, corrected and published by a human editor, who is accountable for what appears. The full account — the role each AI system plays, what the scripts verify, what nothing verifies, and why no AI system is listed as an author — is in AI-USE.md, and appears on the Notice page of both offline editions.

Acknowledgements

arXiv — papers are discovered through arXiv's public APIs, OAI-PMH interface and RSS feeds. Their descriptive metadata is stored, and each page links back to the paper's abstract page rather than hosting the paper. Thank you to arXiv for use of its open access interoperability.
OpenAlex — open-access articles published outside the preprint servers are discovered through the OpenAlex API. Please cite OpenAlex as: Priem, J., Piwowar, H., & Orr, R. (2022). OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv:2205.01833.
Hermes Agent — the agent framework this knowledge base is built and maintained with.
Blume — the documentation framework whose design and layout conventions this front end follows.

License

Code is MIT licensed; the knowledge-base content is released to the public domain under CC0 1.0. See LICENSE.

For Contributors

Tech stack

Astro 7, deployed to GitHub Pages at base path /aied by GitHub Actions. Pagefind provides full-text search; RSS, sitemap and schema.org JSON-LD are generated at build time. Content is markdown in collections defined in src/content.config.ts. The build runs astro check, astro build, Pagefind, then a service-worker step.

Quick start

Requires Node.js 22 or newer and Python 3.11 or newer (the gate scripts use the standard library only). Rebuilding the offline editions additionally needs pandoc, plus the weasyprint and pikepdf Python packages.

git clone https://github.com/edtechdev/aied.git
cd aied
npm install
npm run build     # astro check + astro build + Pagefind + service worker
npm run preview   # serve dist/ locally

That is the whole setup: no environment variables, no API keys, nothing beyond npm install.

Two things to know before editing:

raw/ is not committed. It holds the extracted full text of each source paper, which is gitignored because it is copyrighted material, and every page's sources: field points into it. The site builds without it, but the number-grounding check reports [no raw source found - cannot verify] for an edited article until you fetch that paper's text yourself.
Scope the gates. npm run verify walks the whole corpus, and most of its time goes into the inline-link scan. For an ordinary edit use python3 tooling/scripts/run-gates.py --changed, which reads only the pages you touched. The full sweep is for changes to slugs, facet values or generated artifacts.

Repository layout

├── content/
│   ├── en/            # the source locale: articles, concepts, faqs, resources, newsletters
│   ├── es/            # translations, same collections
│   └── ...            # one folder per locale code in site.config.json
├── raw/papers/        # extracted source text, named by DOI/slug — gitignored
├── pdf-sources/       # source PDFs — gitignored (copyright)
├── src/
│   ├── config/        # siteConfig.ts (wraps site.config.json with types)
│   ├── i18n/          # chrome strings per locale (ui.<code>.ts, pages/*.ts) and
│   │                  # locales.json, the translated per-locale prose
│   ├── layouts/       # BaseLayout.astro (nav, search, footer)
│   ├── lib/           # jsonld.ts, markdown.ts, pageTitles.ts
│   ├── components/    # JsonLd.astro, SourceButtons.astro, per-page components
│   └── pages/         # index, journal, newsletter pages (news/), search, faq, resources, ai, rss, dynamic routes
├── public/            # llms*.txt, the committed EPUB/PDF editions, cover and marks
├── tooling/           # the pipeline: gates, generators, builders, cron prompts and shims
├── skills/            # public-safe mirrors of the agent skills the pipeline follows
├── docs/              # json-ld.md, newsletters.md, and design notes
├── site.config.json   # single source of truth for site-wide metadata
├── wiki.config.yaml   # pipeline config: content paths, gate commands, scan sources, journal feeds
├── concepts.registry.yaml  # the concept vocabulary and redirect map
└── .github/workflows/ # build and deploy to GitHub Pages

Site configuration

site.config.json at the repo root holds every site-wide value: name, short name, brand label, description, language, site URL, base path, repo and issues URLs, editor name and contact, license, PWA theme colors, and the offlineBaseUrl the editions download from. Astro imports it through src/config/siteConfig.ts; the Python tooling opens it with json.load. Edit the config rather than hardcoding a value.

That file is deliberately kept small enough to read. A locale is split in two: site.config.json holds its configuration (code, label, writing direction), and src/i18n/locales.json holds its translated prose — the machine-translation notice, the llms preamble, the offline edition's title and resource labels, and the taxonomy section headings. siteConfig.ts merges the two, so components and the Python builders all read one complete entry. Adding a language means one entry in each file plus its content.

The same file holds the two blocks behind the AI-use disclosure: contributors (the people accountable, referenced by id from page frontmatter so no name is written into a page) and aiDisclosure (the models in use and the harness). The disclosure gate refuses a page that names a model the config does not list, so add a new model to the config first.

Page types

Every page type shares the typed metadata fields and a [[wikilink]] body. They differ in shape:

Article — synthesis blockquote → Key Findings → body sections → What this means for practice → Limitations → Citation last. One page per paper. No Connected section: concept links live inline in the body, and the page's derived view is its metadata table.
Concept — synthesis → Questions to Consider → Introduction → narrative body → Connected Concepts → Connected Articles.
FAQ — frontmatter → question heading → narrative answer. Curated, with no sources and no Citation.
Resource — a fact block (visit, source, author, type, access, license, last verified) → short prose → metadata table → Connected Concepts and Resources. These are not papers, so there is no Citation.
Newsletter issue — see docs/newsletters.md.

audit-article-sections.py enforces the article contract: section order, 3–5 practice bullets, 2–6 limitations bullets, and Citation last.

Typed metadata

The concepts a page touches are named in facet fields — foundations, pedagogy, technology, assessment, methods, stakeholders, institutions, ethics — each taking concept slugs filed under that field's own section of concepts.registry.yaml. Alongside them sit the phrase fields research_method, discipline, level, audience and page_kind. They render as the metadata table at the foot of every page and drive the Pagefind facets and the page's schema.org keywords. A value of the wrong kind fails the build. See tooling/SCHEMA.md; the JSON-LD mapping is in docs/json-ld.md.

Pipeline commands

python3 tooling/scripts/run-gates.py --changed   # the gates, scoped to edited pages
npm run verify                                   # every hard gate, whole corpus
npm run build                                    # the static site
npm run preview                                  # preview the production build
bash tooling/scripts/commit-if-green.sh msg.txt <paths...>   # gate, scan, commit with AI-use trailers
python3 tooling/scripts/generate-llms-files.py   # regenerate llms*.txt (opt-in)
python3 tooling/scripts/validate-facets.py       # check typed metadata against the registry
python3 tooling/scripts/gen-concept-artifacts.py # regenerate concept views from the registry
python3 tooling/scripts/audit-metadata.py --strict
python3 tooling/scripts/check-us-english.py --include-docs
python3 tooling/scripts/sync-skills.py --check   # drift between skills/ and the agent's installed copies

The scheduled jobs run thin shims under <AGENT>/scripts/, and tracked templates for them live in tooling/cron/: aied-scan-prefetch.py (the daily scan's candidate fetch) and aied-weekly-topic-sweep.py (the weekly OpenAlex topic sweep). To install one, copy it into the agent's scripts directory and either export AIED_WIKI or replace the <WIKI> placeholder with your clone path.

The offline editions are built with bash tooling/build-epub.py and committed, so regenerate them after a content change rather than editing the files. Do not run that alongside npm run build: both write into dist/.

How it stays current

Three paths add content, each committing through the same gate route:

Daily scan — searches arXiv (cs.CY, cs.HC, cs.CL, cs.AI, physics.ed-ph) and EdArXiv for new AI-in-education papers, plus a daily OpenAlex journal watch over the outlets this corpus actually draws from.
Weekly journal scan — ingests open-access articles from journal RSS feeds, plus a weekly OpenAlex topic sweep for the outlets nobody watches. The feed list and each journal's freshness window live in wiki.config.yaml under journal_scan.feeds; batch-publishing journals get a longer window so a quiet mailing does not blank them out.
Manual ingestion — PDFs or preprint URLs can be sent at any time.

The two OpenAlex shapes are configured in wiki.config.yaml → scan.sources, each with a cadence (daily or weekly) so a high-volume shape cannot run as part of the cheap pass:

python3 tooling/scripts/cron_prefetch.py --source "journal watch" --days 4   # daily
python3 tooling/scripts/cron_prefetch.py --cadence weekly --days 7          # weekly sweep

The journal watch filters on primary_location.source.id and queries one AI term per request, so a term is not diluted by a multi-word phrase; the topic sweep filters on topics.id and requires a learner or teacher term to keep the list readable. Neither passes --edu-only or --with-pdf-only: the first discards education-technology work that OpenAlex files under information systems or the discipline, and the second discards open-access records that carry is_oa: true with no pdf_url. See tooling/references/openalex.md and the scholarly-discovery-recall skill.

Each run filters for relevance, skips what is already ingested, creates article pages, enriches the concept pages the new work contributes to, rebuilds and deploys. It also records every candidate concept pairing in concept-screen.yaml as integrated or deliberately declined.

Every article ingestion does two things beyond creating the page: it adds inline [[wikilinks]] for every concept the paper touches, and it weaves the finding into each concept page the article contributes to, rather than only adding a Connected Articles entry.

The weekly newsletter

A weekly issue summarizes the most significant research and resources added, plus internal changes to the site. Issues live in content/en/newsletters/<issue-date>.md, render at /news/<issue-date>/, are listed on /news/, and have their own feed at /news/rss.xml. Each issue carries issue (a sequence number) and status (draft, ready or published), so a draft can sit in the folder without appearing on the site.

The format, house style and publishing contract are in docs/newsletters.md: the four section headings, the masthead, the frontmatter, how to link the way the site's own pages link, and the footer.

Three skills drive the work, mirrored under skills/research/:

newsletter-issue-drafting — how to draft an issue: fix the seven-day window, list what was added from git log rather than memory, decide each item's section from the page's own audience: facet, read the page's bottom-line paragraph before writing about it, and verify every link. Its scripts/render-newsletter-preview.py renders a draft to a standalone page in the site palette, so you can review it without rebuilding the site:
```
python3 skills/research/newsletter-issue-drafting/scripts/render-newsletter-preview.py \
  content/en/newsletters/2026-10-24.md --outdir /tmp/newsletter-preview
python3 -m http.server 4322 --bind 0.0.0.0 --directory /tmp/newsletter-preview
```
Serve it on a port other than the site's own preview, which serves the last build.
newsletter-publishing — the issue-as-page contract and the wiring checklist for publishing an issue on a site.
astro-newsletter-publishing — the Astro steps for adding the page family to a site that lacks it: the collection, the routes, the second feed, the icons, the journal and sidebar entries, and the header link.

Troubleshooting

Problem	Check
Site not updating	Confirm the Actions deploy workflow ran: Actions tab → astro-deploy
Search index stale	Search is Pagefind-based — run `npm run build` so `dist/pagefind/` regenerates
llms.txt out of date	`python3 tooling/scripts/generate-llms-files.py`, then `npm run build`
Broken wikilinks	`[[slug]]` must match a file in `content/en/articles/`, `concepts/`, `faqs/` or `resources/`, or a redirect in `src/data/conceptRedirects.ts` / `articleRedirects.ts`
Metadata rejected by the build	`python3 tooling/scripts/validate-facets.py` — a facet value must be a concept filed under that field's own registry section
An article's sections rejected	`python3 tooling/scripts/audit-article-sections.py --changed` — reports missing practice or Limitations sections, bullet counts outside 3–5 and 2–6, and a Citation that is not last
A number is not in its source	`python3 tooling/scripts/verify-number-grounding.py <slug>` — grounded means the token appears in `raw/papers/<file>.md`. Leading-dot p-values, table columns split across lines and values duplicated by HTML conversion are false positives; grep the raw before changing prose
British spelling crept in	`python3 tooling/scripts/check-us-english.py --include-docs` — house style is US English; the checker ignores slugs, link targets and inline code
A resource link has rotted	`python3 tooling/scripts/check-resource-links.py` — re-checks every resource URL and reports the `last_verified` date to bump
YAML parsing errors	Quote titles containing colons: `title: "X: Y"`

AI-use disclosure in practice

Each commit is stamped with AI-Model, AI-Role and AI-Agent trailers, plus Human-Review when a reviewer id is given. Each page carries ai_assist, source_depth, verified and contributors, and reviewed_by only where a person has actually read it. The commit is the per-change record, the frontmatter is the per-page record, and AI-USE.md is the policy. check-ai-disclosure.py enforces valid values; check-ai-disclosure-trailers.py reports commits missing trailers.

Run Your Own Knowledge Base

Everything needed to stand up an equivalent knowledge base for another domain is in tooling/:

tooling/README.md — setup guide
tooling/SKILL.md — the ingestion and export pipeline, written as an agent skill
tooling/SCHEMA.md — page conventions, the typed metadata fields, and the generated vocabulary lists
tooling/scripts/ — the gates and generators: registry and facet checks, the section audit, number grounding, the US-English checker and its fixer, AI-use disclosure, the commit route, the llms generator, the RSS fetchers and the skill sync
tooling/references/ — pipeline architecture, filtering strategies, recovery procedures
tooling/cron/ — the scheduled-job prompt templates
tooling/example/ — starter files

The agent skills the pipeline follows are mirrored in skills/research/. They cover inline linking, article and concept page quality and structure, batch ingestion QA, discovery recall, source acquisition, page deletion, the EPUB/PDF export, site quality, page deepening, the Astro frontend, FAQ and resource pages, citation form, concept-page design and concept-weave screening, deduplication, backlog triage, bulk link editing, answering questions from the corpus and presenting the answers, keeping an already-ingested source current when it is published, web-report-only sources, commit and README hygiene on a public repo, the three newsletter skills, RSS and feed authoring, AI-content provenance and public-repo PII hygiene. scholarly-discovery-recall is the one to read when the scan is missing papers the maintainer finds by hand: it carries the measurement procedure and the OpenAlex API traps, and tooling/references/openalex.md holds the invocations.

Those mirrors are scrubbed, public-safe copies of the agent's own installed skills: the repo copy uses <WIKI> for the repository root and <AGENT> for the skill store, and names no person and no agent product. Everything else is identical. python3 tooling/scripts/sync-skills.py --check normalizes those sanctioned differences and reports only real content drift; --to-repo refreshes a mirror with the redaction applied, and --to-installed copies the other way. Which side is authoritative depends on the file: a skill's SKILL.md is behavior, so the installed copy wins, while tooling/scripts/ is the working copy the pipeline runs, so the repo wins.

AGENTS.md documents the page-structure rules and the inline-link requirement an agent must follow on every ingestion.

To start your own, copy tooling/ and skills/ into a new repository and follow tooling/README.md. No API keys are required.

Navigation Menu

Repository navigation

About

Repository files navigation

AI in Education Knowledge Base

For Readers

What's inside

Use it with your own AI assistant

Translations

How this is made

Acknowledgements

License

For Contributors

Tech stack

Quick start

Repository layout

Site configuration

Page types

Typed metadata

Pipeline commands

How it stays current

The weekly newsletter

Troubleshooting

AI-use disclosure in practice

Run Your Own Knowledge Base

About

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
content		content
docs		docs
public		public
skills		skills
src		src
tooling		tooling
.gitignore		.gitignore
.nojekyll		.nojekyll
AGENTS.md		AGENTS.md
AI-USE.md		AI-USE.md
AIED-BACKLOG.md		AIED-BACKLOG.md
LICENSE		LICENSE
README.md		README.md
astro.config.mjs		astro.config.mjs
concepts.registry.yaml		concepts.registry.yaml
generate-sw.mjs		generate-sw.mjs
package-lock.json		package-lock.json
package.json		package.json
pagefind.yml		pagefind.yml
resource-wiring.yaml		resource-wiring.yaml
site.config.json		site.config.json
tsconfig.json		tsconfig.json
wiki.config.yaml		wiki.config.yaml

Latest commit

History

Folders and files

Repository files navigation

AI in Education Knowledge Base

For Readers

What's inside

Use it with your own AI assistant

Translations

How this is made

Acknowledgements

License

For Contributors

Tech stack

Quick start

Repository layout

Site configuration

Page types

Typed metadata

Pipeline commands

How it stays current

The weekly newsletter

Troubleshooting

AI-use disclosure in practice

Run Your Own Knowledge Base

About

Resources

Stars

Watchers

Forks

Contributors

Languages