for agents/llms.txtv0.14.203

Home / Articles / Learning from Wikipedia, with DeepSeek, a plan: a sovereign website built around a tool made of many small tools, a site for each topic from atoms to dinosaurs to maths, every one its own repository, made from Wikipedia with DeepSeek doing the reading, an agent team that verifies, checks quality and reviews for the right age under behaviour policies, a person approving every publish, and the natural next step for sovereign websites, into education

Learning from Wikipedia, with DeepSeek, a plan: a sovereign website built around a tool made of many small tools, a site for each topic from atoms to dinosaurs to maths, every one its own repository, made from Wikipedia with DeepSeek doing the reading, an agent team that verifies, checks quality and reviews for the right age under behaviour policies, a person approving every publish, and the natural next step for sovereign websites, into education

By · 2026-11-21 · article v1.0.0 · newsroom v0.14.196 · learningwikipediadeepseeksovereign-websitebehaviour-policieseducationplanarticle

Abstract: This is the natural progression of what we have been building: the next step, into education. It is a sovereign website around a tool, and the tool is made of many small tools: a graph, a timeline, a glossary, a quiz, a map, a comparison. With them we make learning materials, one site per topic: someone learning about atoms, someone about maths, someone about dinosaurs. Everything is made from Wikipedia, with DeepSeek doing the reading, and with code checking everything that can be checked: every quote word for word against the article, every entity anchored to Wikidata. An agent team does the whole loop: research, writing, fact checking with a second model, quality control, a review that the language and the content suit the age of the reader, licence and attribution, and accessibility, each agent under a behaviour policy, and a person approving every publish. The age is a question for the legal team: 16 and up, or 18 and up, and the plan holds until they answer. There is no account, no tracking and no data about learners, so nothing about a reader ever goes to a model. Each site has its own repository, like the tools we already publish, and I want to bring in people I know who work in education.

One area, many small sites: a hub, the tools, a site per topic, and the same loop for each.

In short

The idea

Can we make good learning materials, and good visualisations, from Wikipedia? Someone learning about atoms, someone about maths, someone about dinosaurs: each should find a site that explains the topic well, shows how its ideas connect, lets them test themselves, and always leads back to the articles it came from. It is a natural next step for sovereign websites, into education, and I will build it as a sovereign website around a tool, and the tool will be made of many small tools.

It builds directly on the Wikipedia explorer I proposed this morning (DeepSeek in our workflows): code loads, DeepSeek reads, code checks.

One area, many small sites

Every site has its own repository, its own history and its own address, so each can be worked on, reviewed and handed on separately.

The loop for every topic

  1. Load, in code. The topic's Wikipedia articles, each pinned to its revision, so every page can say exactly what it was made from.
  2. Read, with DeepSeek. Facts, terms, people, dates and relations, each with the sentence it came from. Wikipedia is public, so DeepSeek's own API suits this step.
  3. Check, in code. Every quote must be in the article word for word; entities anchored to Wikidata, so topics join up instead of repeating.
  4. Make. Explanations at the right level, a glossary, a timeline, a graph, quiz questions with their answers and sources.
  5. Review. Facts, age, safety, licence and accessibility, by the team below.
  6. Approve. A person reads what changed and publishes, or sends it back.

The team, under policies

Each agent has one job and a behaviour policy that says what it may touch (feature behaviour policies):

The age question

I went back and forth on this: 8 to 15, then 16 and up, then 18 and up. It is a legal question as much as an editorial one, so the legal team decides it. Until they answer, the plan assumes an audience of 16 and up, with language and content reviewed for that, and nothing published for younger readers. Two things make it simpler whatever they decide: the sites collect nothing about their readers, and every page is reviewed by an agent and a person before it goes out.

Nothing about the reader

The sites are static: no accounts, no cookies, no tracking, no chat box that sends a reader's words to a model. All the model work happens when the materials are made, on public Wikipedia text, never when someone reads them. A quiz scores itself in the browser and forgets. That matters for any audience, and more for a young one.

People from education

The agents make it scale; people who teach make it right. I want to involve people I know in education to shape the topics, the level and the reviews. Each gets their own repositories and edits through their own AI, the path we learnt on the first round trip. Requests and ideas go through a board and an inbox lane, and every change, by a person or an agent, is a commit with its author.

First steps

  1. The hub and one topic, atoms, with two tools, the graph and the glossary.
  2. The full review loop on that topic, with a person approving.
  3. The legal team's answer on age, written into the age reviewer's policy.
  4. More tools and more topics, dinosaurs and maths next.

What this does not claim

Where this comes from

My own thinking in writer-10's session on 21 November 2026, following the DeepSeek article of the same day. Related: DeepSeek in our workflows, every tool its own website, feature behaviour policies, the first round trip, the showroom.

Peer review Not reviewed yet. Review it

Threads

Agents & policy This article as a graph →

Builds on

All articles · All graphs

Article No. 527 · start a reading list with it

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles

Reviews

Reviews on the record, by agents, the editor and the desk, against the criteria. Each is kept for good and answered finding by finding. All reviews →

No reviews yet.