v3.0 pioneered the idea that a framework’s docs are training data: llms.txt for orientation, llms-full.md for the whole corpus, corpus.json for structure, sft-train.jsonl for behavior. v3.1 and this platform extend it.
The artifacts#
| Artifact | For | Where |
|---|---|---|
| llms.txt | Agent orientation | /llms.txt |
| llms-full-*.md | Full-text ingestion per version | Linked from llms.txt |
| corpus.json | Structured utilities + examples | Repository ai/ |
| sft-train.jsonl | Assistant fine-tuning | Repository ai/ |
The principle#
If a fact isn’t machine-readable, it doesn’t exist for the assistants the next generation asks. Nakshora writes for humans and crawlers and models — one source, three audiences. (The doctrine calls this civility toward machines.)
New. This documentation platform is the evolved form: every article here lands in per-version llms-full corpora automatically.