# llms-deep-optimizer Audits and rewrites any llms.txt, llms-full.txt, llms-small.txt, or llms-facts.txt until it passes a measurable bar, inside a multi-pass convergence loop with live probes. An llms file is a **promise list** — every link and every fact must pay off — not a document to make read well. `llms-deep-optimizer` treats it that way: detect what the file is, run every pass (structure, links, descriptions, size ladder, full-file grammar, facts anchors and truth, provenance, serving headers), collect every finding before writing anything, fix everything Medium-severity or above inside a convergence loop, then verify with live keyword, vector, and agent-usability probes. It also builds a **topical llms file from scratch** given a pool of uncategorized facts — the same compile step [notes-to-llms-txt](/skills/notes-to-llms-txt/) hands its draft to once a first structure exists. Try a bounded, single-pass version of the audit live on the [optimizer playground page](/playground/optimizer/) — paste an existing llms.txt and see one pass's worth of findings and fixes. The real skill runs the full multi-iteration convergence loop with live link/probe verification; the playground demo shows the technique, not the whole loop. **Use it for:** "optimize this llms.txt", "why does my llms-full break", "check the facts file", "build an llms file for X", "make this docset navigable" — any file named `llms*.txt` or a `.llms/` directory. **Not for:** a skill file or hub spoke (`skill-optimizer`) · a prompt (`prompt-deep-optimizer`) · ordinary prose (`document-deep-optimizer`) · crawling a site (`web-text-mirror`) · drafting a first structure out of disorganized notes — that's this skill's own upstream step ([notes-to-llms-txt](/skills/notes-to-llms-txt/)).