One folder that teaches any model how to explain any topic as a single, white, research-paper-style HTML page: centered column, serif text, tiny monospace labels, amber highlights on the sentences worth remembering, CSS-drawn diagrams, a concept map, and a references list where every entry says what it supports.

The knowledge lives in an Open Knowledge Format (OKF) bundle: plain Markdown with YAML frontmatter, no SDK, any model can read it.

philosopher-okf/
├── SKILL.md                  # entry point for skill-aware agents (Claude, Codex, Cursor)
├── okf/                      # the OKF bundle
│   ├── index.md              # root index (declares okf_version 0.2)
│   ├── log.md                # change history
│   ├── core/                 # workflow, classifier, contract, template, writing, sources, depth, graph
│   └── domains/              # 12 study domains, one file each
├── scripts/validate.sh       # OKF conformance + link check
└── examples/
    └── sample-flamingo-reading-guide.html   # what the output looks like

Technology and engineering · Medical and health · Natural sciences · Mathematics, statistics and logic · Finance and accounting · Economics · Business and management · Law and policy · Humanities, philosophy and history · Social sciences and psychology · Research papers · General

1. Reads okf/core/workflow.md (the 8-step procedure, lazy load table, token budget).
2. Classifies the question (core/classifier.md) into one primary domain.
3. Loads only what the depth needs: brief = template + the domain's # Quick card (~3 reads), standard adds source policy and depth policy, deep adds the graph method and the full domain file. Research is capped per depth (brief 2 searches / 0 fetches, standard 4 / 2, deep 8 / 4), and fetched pages are noted, not re-read.
4. Writes <topic>.html in one call. The user opens it in a browser.

Editors and chat views each ask for one thing — a folder, a link, or plain text — so the skill ships in all three shapes. A skill agent opens SKILL.md and walks the bundle from there; a browsing model needs only the link to philosopher.md, which contains the whole skill in one file and does the rest on its own; anything without network is handed the prompt below. The okf/ folder is the source of truth: philosopher.md is a generated single-file copy, rebuilt by GitHub Actions whenever okf/ changes — never edit it by hand (bash scripts/build-single.sh rebuilds it locally).

No terminal at all? Open the site, click copy skill, and paste it into any chat — you're done.

One paste covers every chat view — copy this prompt:

You are Philosopher, a teaching engine. Open https://github.com/frypan05/philosopher-OKF/blob/main/philosopher.md and follow it exactly — it contains everything: the workflow, the page template, and the domain cards. If you cannot open URLs, ask me to paste philosopher.md. Reply with the file path only, or one HTML code block if you cannot create files.
My question: <your topic>

Chat apps that cannot create files print the HTML in one code block — save it as <topic>.html and open it in a browser.

Ask in plain words, optionally with a level and depth: "Explain how vaccines train the immune system, for a 9th grader, brief", "Explain this paper: https://arxiv.org/abs/2204.14198".

bash scripts/validate.sh okf

Checks the three OKF rules (frontmatter, non-empty type, reserved files), that index.md has no frontmatter except the root's okf_version, that log.md uses ISO dates, and warns about broken links.

1. Add okf/domains/<slug>.md with type: Study Domain (copy any existing domain).
2. Add a row to the routing table in okf/core/classifier.md.
3. Add an entry in okf/index.md and okf/domains/index.md.
4. Add a dated line to okf/log.md, then run the validator.

core/classifier.md routes by keywords and intent. If your pipeline has its own classifier, pass its label as one of the domain slugs and the skill uses it directly (rule 1 in the classifier file).

• Token targets per run (whole context, including research and the written page): brief under 12k, standard under 25k, deep under 45k. The levers are the workflow's research caps, the lazy load table, domain Quick cards, and writing the HTML in a single call. A run that skips fetched pages in favour of notes will land well under these numbers; a run that pulls several full documents will not.
• Source lists point to host-level addresses of well-established sites. They were not link-checked from the build environment, so run a link checker in CI before publishing.
• The sample page was built from the paper's public abstract record and the section list, and is marked as such. Check figures against the full text.
• Models without web access can only cite from memory. The source policy makes them say so on the page instead of inventing links.

MIT. See LICENSE.