brain-process-inbox

name	brain-process-inbox
description	Process new raw notes, documents, and thoughts from month folders under 00-inbox into the personal knowledge base.

Process Inbox Skill

Follow AGENTS.md first. This skill only defines the task-specific inbox procedure.

Goal

Convert raw captured material into durable, linked knowledge with minimal human effort.

Inputs

• New or modified files under 00-inbox/YYYY-MM/

Outputs

• Organized editable notes and source-derived material under 10-notes/
• Updated canonical notes in 20-knowledge/ or active project contexts in 30-projects/
• Updated person profiles in 20-knowledge/people/ when captures contain stable people facts, preferences, needs, communication patterns, relationship or other similar context
• Updated indexes
• Updated log
• Updated unresolved/duplicate/stale-claim reports if needed

Steps

1. Scan 00-inbox/YYYY-MM/ month folders.
2. Classify each item:
   • domain
   • type
   • provenance
   • sensitivity
   • confidence
   • likely destination
3. Extract useful material from each item:
   • durable facts
   • source context
   • decisions or preferences
   • open questions
   • possible duplicate/overlap signals
   • stale, contradictory, or time-sensitive claims
   • proposed canonical note updates
   • affected people profiles and whether they need concise updates
4. Create or update organized material in 10-notes/ by preserving the captured content nearly as-is, except for meeting/call/recording/transcript captures.
   • Use short descriptive kebab-case filenames.
   • Do not add leading date prefixes to source filenames.
   • Preserve the human's wording, examples, and context as much as practical.
   • Light cleanup is allowed: structure, formatting, minor spelling fixes, broken Markdown, metadata, backlinks, and short agent notes that flag clear issues.
   • Do not rewrite captured content into LLM prose and do not replace it with a summary.
   • For meeting/call/recording/transcript captures, follow the meeting compression rules below instead of copying the full raw record into 10-notes/.
   • Preserve capture/import dates in frontmatter, inbox processed markers, and the knowledge log.
5. Search existing canonical notes before creating new pages.
6. Update existing notes when possible.
7. Create new notes only when needed.
8. Add frontmatter.
9. Add backlinks.
10. Update 20-knowledge/index.md when navigation changes.
11. Append a log entry to 20-knowledge/log.md.
12. Update .generated/indexes/.
13. Add unresolved questions to .generated/unresolved.md.
14. Add possible duplicates to .generated/duplicates.md.
15. Add stale or contradictory claims to .generated/stale-claims.md.
16. Produce a concise summary of changes.

Skill-specific judgment

• Keep the inbox easy for the human: do not require extra categorization before processing.
• Prefer updating existing knowledge/project/system notes over creating new pages.
• If destination is unclear after a reasonable search, add the item to .generated/unresolved.md.
• Write canonical notes in the original language of the source material. If a source mixes languages, follow the dominant language and preserve key quoted terms as written.

Meeting compression judgment

• Leave full meeting/call/recording/transcript records in 00-inbox/YYYY-MM/; do not copy them verbatim into 10-notes/.
• Create one faithful compressed meeting note in 10-notes/ or the clearly implied project location.
• Preserve substantive facts, decisions, nuances, context, assumptions, risks, constraints, open questions, commitments, useful side topics, etc.
• Remove greetings, repetition, filler, false starts, meaningless self-corrections, over-explaining, etc.
• Let sections emerge from the meeting topics; do not force a fixed template.
• Consolidate scattered decisions and action items without losing context.
• Use light speaker/timestamp refs only when useful for verifying important nuance, decisions, ambiguity, or high-impact facts.
• Link to the raw inbox file with a quoted Obsidian wikilink; do not use traversal paths.
• Keep sensitivity at least as high as the raw source.

Metadata judgment

• Use origin: external for digests or notes from someone else's article, book, course, lecture, video, recipe, official record, vendor page, professional guidance, or technical reference.
• Use origin: personal for the human's own memories, reflections, beliefs, preferences, original work, personal records, practices, planning, or decisions.
• Use origin: mixed when personal context or interpretation and external material both matter.
• Use origin_note for a short provenance explanation, such as "personal memory and preference", "digest of an external article", or "personal interpretation of external research".
• Use sensitivity: high for health, family, finance, legal, identity, personal documents, private work/client context, or anything harmful if exposed.
• Use sensitivity: medium for personal preferences, non-public work methods, reusable prompts with business relevance, and private but low-risk research notes.
• Use sensitivity: low for public concepts, public tools, public articles, and general learning notes.
• Use confidence: high for direct, reliable evidence.
• Use confidence: medium for plausible interpretation.
• Use confidence: low for memory, hypotheses, rough notes, or AI inference.
• Do not copy high-sensitivity content into low-sensitivity summaries.
• Write metadata file references as quoted Obsidian wikilinks.