convert-note

Convert Note — Raw-to-Atomic Decompiler

You are ingesting raw content from the raw/ folder into the Infinite Brain knowledge graph.

Steps

1. Ask: "Which file in raw/ should I convert? (or 'all' to process everything)"
   • Source files need no pre-organization — a descriptive filename (e.g. 2026-05-15-karpathy-llm-wiki.md) is enough; tagging and classification happen here, not in raw/.
2. Scan the target file(s) with Read. Treat source files as immutable — do not modify them.
3. Decompose into atomic nodes — one concept per node, 50-300 words each. Never exceed 300 words: longer nodes defeat scoped retrieval.
4. For each node:
   • Classify with exactly one of the 16 content types: pillar decision concept question playbook task event pattern hypothesis fact source bookmark note contact reference custom
   • Assign a unique id in type-descriptive-slug format (kebab-case)
   • Populate all frontmatter fields from _system/FRONTMATTER-SCHEMA.md
   • Wire to at least one other node using the 10 edge types: related_to depends_on derived_from contradicts supports part_of preceded_by followed_by authored_by tagged_with
   • Place in the folder matching the type — the plural of the type name (pillar → pillars/<id>.md, concept → concepts/<id>.md, …; irregular: hypothesis → hypotheses/)
5. Write each node file.
6. Mandatory: Append each new node's row — id, summary, edge count, matching the | ID | Summary | Edges | table columns — to _system/INDEX.md under the correct type section. If the index is not updated, /query-vault's scoped retrieval fails and the vault goes blind.
7. Move the processed source file from raw/ to raw/processed/ (e.g., raw/article.md → raw/processed/article.md).
8. Write a log node to logs/log-convert-note-YYYYMMDD-HHmmss.md using the log schema from _system/FRONTMATTER-SCHEMA.md:
   • operation: convert-note
   • affected_nodes: list of all node IDs created
   • summary: one sentence — source file name + node count
   • Body (30–80 words): what was processed, what nodes were created, any notable decisions made during decomposition.
9. Confirm: "Converted X nodes from raw/filename. Original moved to raw/processed/. Log written."

Rules

• Never merge concepts — err toward more atomic nodes.
• summary must be 1-2 sentences under 200 chars.
• confidence must reflect actual certainty from the source.
• visibility defaults to namespace unless clearly indicated otherwise.
• staleness_signal must be a specific observable condition.
• tags: 2-8 per node, lowercase kebab-case.
• namespace: infer from the content domain.
• Do NOT create nodes with type: raw — raw/ is only an inbox.
• Always check _system/INDEX.md to avoid duplicate IDs before writing.
• Write title, summary, body prose, and edge note fields in the vault's document language (see _system/AGENTS.md → Document Language); keep id slugs, type, edge type names, visibility, namespace, and tags in canonical English. Note content may be in a different language than the source file — convert it.