Okf ingest

Ingest a source or new concept into an Open Knowledge Format (OKF) knowledge base (a wiki/ bundle of Markdown + YAML frontmatter). Use when the user wants to add a paper, article, URL, video, repo, dataset, person, project, or note to their wiki / knowledge base / second brain. Triggers: "ingest this", "add this paper/article to the wiki", "capture this source", "save this to my KB", "add a note about X". Fetches and VERIFIES the source, writes a conformant page, cross-links it, updates the index + changelog, and lints.From its SKILL.md

Install

npx -y skills add thattimc/skills --skill okf-ingest

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

3 things to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
runs commandsInstructs the agent to run 2 commands, including `curl -sL "http://export.arxiv.org/api/query?id_list=<id>"` and 1 more.
fetches URLsInstructs the agent to fetch 2 URLs, including http://export.arxiv.org/api/query?id_list=<id> and 1 more.

SKILL.md

3.4 KB, 792 tokens by cl100k_base, as published. Nobody here has run it

Ingest into an OKF knowledge base

Add a source or concept to an OKF wiki as one Markdown file (a "concept"). If the repo has a CLAUDE.md, that is authoritative — follow it (this skill summarizes the same rules so it works standalone).

0. Orient

The bundle is the wiki/ directory. Edit knowledge only there. A concept's id is its path minus .md (wiki/papers/foo.md → papers/foo).
If tools/okf_lint.py exists, you'll run it at the end.

1. Verify — never fabricate

Fetch the real source and confirm its facts from the source itself, not memory (your training cutoff may predate it):

arXiv: curl -sL "http://export.arxiv.org/api/query?id_list=<id>" — read the canonical <title>, <author> list, and <published> date. If the id is unknown, search: ...query?search_query=ti:%22<title words>%22&max_results=5.
Web: use WebFetch on the URL; capture title, author/site, date.
If you cannot confirm it exists, stop and tell the user — do not invent metadata.
Optionally save the raw capture under raw/ (read-only provenance, outside the bundle).

2. Place it

Choose a domain folder by subject (wiki/papers/, wiki/people/, …) — folder = subject, type = kind. Create the folder + an index.md if it's new.
Choose a lowercase, hyphenated slug (matches [A-Za-z0-9_][A-Za-z0-9_.-]*, no spaces).

3. Write `wiki/<domain>/<slug>.md`

Frontmatter (keys in this order; use templates/<type>.md as a stamp if present):

---
type: Source            # Source | Entity | Concept | Note | Overview | Reference
resource: <URI>         # required for Source
title: <human title>
description: <one sentence>
tags: [a, b]            # a YAML LIST, never a comma string
timestamp: 2026-01-01T00:00:00Z   # UTC ISO 8601; refresh on every edit
# extensions allowed after the standard keys (authors, year, venue, doi, …)
---

Body: grounded summary prose, then optional # Key contributions / # Schema / # Examples (fenced code is fine), then # Citations (the source's own resource URI first). Ground every claim in the source.

4. Link, index, log

Add genuine relative markdown cross-links to/from related concepts — [title](../other/x.md). No [[wikilinks]], no leading /.
Update the domain index.md (* [Title](slug.md) - description) and the root wiki/index.md if you added a new domain.
Prepend a wiki/log.md entry under today's date (* **Creation**: Added [Title](path).).

5. Lint

python3 tools/okf_lint.py     # fix any errors; review warnings

Scaling up

For many sources, or anything post-dating your knowledge cutoff, fan out with a workflow: discover → verify each against arXiv/web → author. Verification-before-authoring is what keeps the KB free of hallucinated sources.

What ships with it: 1 file

109 B alongside SKILL.md

agents/

openai.yaml109 B