notes-to-llms-txt

Raw notes are multi-topic, unordered, and mix genres — a fact next to a todo next to a half-formed question next to a name. notes-to-llms-txt sits upstream of both llms-concept-abstractor (which assumes you already know the one concept to pull out) and llms-deep-optimizer (which assumes structure already exists to audit): it takes the mess as-is, finds the topics actually in it, and drafts a first llms.txt family good enough for the optimizer to take the rest of the way.

How it works

  1. Ingest every source in full — no sampling.
  2. Segment into atomic units (bullet, paragraph, heading-scoped block), tagging each unit’s genre: fact, todo, decision, question, name/entity mention, or noise.
  3. Cluster the surviving units by topic — notes are rarely about one thing, so clustering goes by subject, not by which file or meeting a unit came from.
  4. Draft an entry per topic with enough content: title, description, source-anchored facts, verbatim links, open questions left as questions rather than smoothed into facts.
  5. Compile and hand off to llms-deep-optimizer for the real multi-pass audit.

Every fact traces to a specific note; a gap in the notes is a gap in the output, never filled from general knowledge. Anything secret-shaped — API keys, passwords, connection strings pasted into a meeting note in passing — is redacted before it reaches any output file, since these output families are exactly the kind of thing that ends up pasted into a shared repo or a public library later.

Try a bounded, single-pass version live on the notes-to-llms playground page.

Use it for: “turn my notes into an llms.txt”, “structure these meeting notes”, “my notes are a mess, make them navigable”.

Not for: notes already scoped to one concept (llms-concept-abstractor) · an llms.txt that already exists and just needs auditing (llms-deep-optimizer) · a request that needs new research, not organizing what’s already written down (/dr).