llms.txt optimizer
An llms file is a promise list: every link and every fact must pay off. It is not a document to make read well, which is why the ordinary prose-editing instinct makes these files worse.llms-deep-optimizer audits one against sixteen passes and rewrites it under a demotion guard — no link lost, no fact dropped without saying so.
This is a bounded two-pass demo: one audit pass, one fix pass, 4000 characters of input. The full skill loops audit → fix → re-audit to a convergence exit, runs live link and retrieval probes, and ends with a blind re-audit by a fresh reader. Two passes show the shape of the thing; they do not clear the bar.
Findings
Diff
The rewrite could not be separated from the findings, so the raw output is below instead.
Rewritten
What the audit pass looks for
| Pass | Judges |
|---|---|
| P0 | Which kind of file this is — index, full, small, or facts |
| P1 | Structure: H1, blockquote summary, section headings, link-list shape |
| P2 | Links: malformed, relative where absolute is needed, obviously dead |
| P3 | Descriptions: missing, restated from the title, or uninformative |
| P5 | Size ladder: an index doing a full file's job |
| P7 | Facts shape: are fact lines atomic and anchored? |
| P9 | Provenance: unsourced claims, steering spans, secrets |
High means a reader is misled or blocked — a dead link, an unsourced fact, a steering span, a secret, an index that is really a full file. Medium means a reader pays extra. Steering spans are deleted, never rephrased; secrets are redacted, never described.