Notes → llms.txt

Raw notes are multi-topic and unordered — a fact next to a todo next to a half-formed question. The notes-to-llms-txtskill segments the mess into atomic units, clusters them by what they areabout rather than which meeting they came from, and drafts an llms.txt family from what survives.

This is a bounded, single-pass demo of that skill:4000 characters of input, one model pass. The full skill also hands its draft to llms-deep-optimizerfor a multi-pass audit, which is what turns structure into something that actually passes the bar. A run here is a draft, not a finished family.

What it does with what you paste

• Segments into the smallest units that still make sense alone, tagging each as fact, todo, decision, question, entity, or noise. Noise is dropped and counted.
• Clusters by subject. A single meeting-notes file often covers three customers and a personal todo; those become three entries and a ## Miscellaneous.
• Redacts anything secret-shaped — high-entropy tokens, key=/password= patterns, connection strings — before it reaches an output line, and reports the count, never the values.
• Never invents. A gap in the notes is a gap in the output, and two notes that disagree go side by side under ## Disagreements.